row_id string | section string | section_slug string | resource_type string | marker string | title string | url string | url_kind string | domain string | annotation string | description string | key_contribution string | novelty string | impact string | signal string | signal_strength string | source_readme string | source_line int64 | source_url string | date_added string | collection string | collection_slug string | user_goal string | lifecycle_stages string | audience string | loop_layer string | scope_fit string | evidence_class string | evidence_tier string | source_status string | canonical_url string | source_title string | source_description string | authors string | publication_date string | publication_year string | publication_venue string | publisher string | doi string | publication_note string | primary_category string | metadata_source string | github_repo string | github_stars string | github_forks string | github_license string | github_created_at string | github_updated_at string | arxiv_id string | audited_at timestamp[ms] |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ale-0801 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | EvoCode-Bench: Evaluating Coding Agents in Multi-Turn Iterative Interactions | https://arxiv.org/abs/2605.24110 | external | arxiv.org | Benchmark of 26 evolving coding tasks across 227 evaluation rounds using cumulative executable tests to check that agents keep prior requirements working as specifications change, with top agents reaching only about 50% on multi-turn success metrics. | Benchmark of 26 evolving coding tasks across 227 evaluation rounds using cumulative executable tests to check that agents keep prior requirements working as specifications change, with top agents reaching only about 50% on multi-turn success metrics. | Benchmark of 26 evolving coding tasks across 227 evaluation rounds using cumulative executable tests to check that agents keep prior requirements working as specifications change, with top agents reaching only about 50% on multi-turn success metrics. | Evaluation data is used as the feedback signal for improving loop behavior. Benchmark of 26 evolving coding tasks across 227 evaluation rounds using cumulative executable tests to check that agents keep prior requirements working as specifications change, with top agents reaching only about 50% on multi-turn success me... | Use EvoCode-Bench: Evaluating Coding Agents in Multi-Turn Iterative Interactions to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,508 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1508 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2605.24110 | [2605.24110] EvoCode-Bench: Evaluating Coding Agents in Multi-Turn Iterative Interactions | Coding agents are increasingly used as iterative development partners, but most benchmarks still evaluate one specification followed by one final assessment. This leaves out a basic question: can an agent keep its own codebase working as requirements change? We introduce EvoCode-Bench, a benchmark of 26 stateful coding... | Haiyang Shen; Xuanzhong Chen; Wendong Xu; Yun Ma; Liang Chen; Kuan Li | 2026-05-22 | 2026 | arXiv | arXiv | Work in Progress; 32 pages, 10 figures, preprint | cs.AI | arxiv-api | 2605.24110 | 2026-09-26T17:25:52 | ||||||||
ale-0802 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | On the Reliability of Computer Use Agents | https://arxiv.org/abs/2604.17849 | external | arxiv.org | Repeated-execution study on OSWorld decomposing why computer-use agents fail tasks they previously completed, separating execution stochasticity, task-specification ambiguity, and behavioral variability as distinct causes of unreliability. | Repeated-execution study on OSWorld decomposing why computer-use agents fail tasks they previously completed, separating execution stochasticity, task-specification ambiguity, and behavioral variability as distinct causes of unreliability. | Repeated-execution study on OSWorld decomposing why computer-use agents fail tasks they previously completed, separating execution stochasticity, task-specification ambiguity, and behavioral variability as distinct causes of unreliability. | Links loop design to measurable tasks where progress and failure can be compared. Repeated-execution study on OSWorld decomposing why computer-use agents fail tasks they previously completed, separating execution stochasticity, task-specification ambiguity, and behavioral variability as distinct causes of unreliability... | Use On the Reliability of Computer Use Agents to measure progress and gate completion with repeatable evidence. | Research source arXiv:2604.17849; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,509 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1509 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2604.17849 | [2604.17849] On the Reliability of Computer Use Agents | Computer-use agents have rapidly improved on real-world tasks such as web navigation, desktop automation, and software interaction, in some cases surpassing human performance. Yet even when the task and model are unchanged, an agent that succeeds once may fail on a repeated execution of the same task. This raises a fun... | Gonzalo Gonzalez-Pumariega; Saaket Agashe; Jiachen Yang; Ang Li; Xin Eric Wang | 2026-04-20 | 2026 | arXiv | arXiv | 33 pages, 3 figures, 4 tables | cs.AI | arxiv-api | 2604.17849 | 2026-09-26T17:25:52 | ||||||||
ale-0803 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | AgentLens: Revealing the Lucky Pass Problem in SWE-Agent Evaluation | https://arxiv.org/abs/2605.12925 | external | arxiv.org | Grades 2,614 SWE-agent trajectories across eight models to show that 10.7% of passing trajectories in its 1,815-trajectory evaluation subset are lucky trial-and-error successes, replacing binary pass/fail with process-quality tiers that shift model rankings. | Grades 2,614 SWE-agent trajectories across eight models to show that 10.7% of passing trajectories in its 1,815-trajectory evaluation subset are lucky trial-and-error successes, replacing binary pass/fail with process-quality tiers that shift model rankings. | Grades 2,614 SWE-agent trajectories across eight models to show that 10.7% of passing trajectories in its 1,815-trajectory evaluation subset are lucky trial-and-error successes, replacing binary pass/fail with process-quality tiers that shift model rankings. | Evaluation data is used as the feedback signal for improving loop behavior. Grades 2,614 SWE-agent trajectories across eight models to show that 10.7% of passing trajectories in its 1,815-trajectory evaluation subset are lucky trial-and-error successes, replacing binary pass/fail with process-quality tiers that shift m... | Use AgentLens: Revealing the Lucky Pass Problem in SWE-Agent Evaluation to measure progress and gate completion with repeatable evidence. | Research source arXiv:2605.12925; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,510 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1510 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2605.12925 | [2605.12925] AgentLens: Revealing The Lucky Pass Problem in SWE-Agent Evaluation | Evaluation of software engineering (SWE) agents is dominated by a binary signal: whether the final patch passes the tests. This outcome-only view treats a principled solution and a chaotic trial-and-error process as equivalent. We show that this equivalence is empirically false. We evaluate 2,614 OpenHands trajectories... | Priyam Sahoo; Gaurav Mittal; Xiaomin Li; Shengjie Ma; Benjamin Steenhoek; Pingping Lin; Yu Hu | 2026-05-13 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2605.12925 | 2026-09-26T17:25:52 | |||||||||
ale-0804 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | ORLoopBench: Solver-in-the-Loop Benchmarks for Self-Correction | https://arxiv.org/abs/2601.21008 | external | arxiv.org | Formalizes infeasible-model debugging as a solver-in-the-loop process where each action triggers solver re-execution and infeasibility recomputation, giving deterministic verification for iterative repair in operations research. | Formalizes infeasible-model debugging as a solver-in-the-loop process where each action triggers solver re-execution and infeasibility recomputation, giving deterministic verification for iterative repair in operations research. | Formalizes infeasible-model debugging as a solver-in-the-loop process where each action triggers solver re-execution and infeasibility recomputation, giving deterministic verification for iterative repair in operations research. | Verification is promoted from a final check to a loop-control signal. Formalizes infeasible-model debugging as a solver-in-the-loop process where each action triggers solver re-execution and infeasibility recomputation, giving deterministic verification for iterative repair in operations research. | Use ORLoopBench: Solver-in-the-Loop Benchmarks for Self-Correction to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,511 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1511 | Verify | verify | Gate progress with tests, evals, and evidence. | trigger;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://openreview.net/pdf/16a0193aa4e71ffe6c921ac0081a66b525eea017.pdf | [2601.21008] ORLoopBench: Solver-in-the-Loop Benchmarks for Self-Correction and Behavioral Rationality in Operations Research | Operations Research practitioners debug infeasible models through an iterative process: inspecting Irreducible Infeasible Subsystems ( IIS), identifying constraint conflicts, and repairing formulations until feasibility is restored. Existing LLM benchmarks mostly treat OR as one-shot translation from problem descriptio... | Ruicheng Ao; David Simchi-Levi; Xinshang Wang | 2026 | 2026 | Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306 | PMLR | Published in Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306; the linked arXiv record remains available for open access. | cs.LG | PMLR camera-ready record | 2601.21008 | 2026-09-26T17:25:52 | ||||||||
ale-0805 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis | https://arxiv.org/abs/2605.30434 | external | arxiv.org | Benchmark of 68 real-world data-analysis tasks built from Kaggle notebooks spanning 2,225 interactive turns, finding that long-horizon errors account for 52-69% of agent failures and that maintaining a correct analytical state is the core bottleneck. | Benchmark of 68 real-world data-analysis tasks built from Kaggle notebooks spanning 2,225 interactive turns, finding that long-horizon errors account for 52-69% of agent failures and that maintaining a correct analytical state is the core bottleneck. | Benchmark of 68 real-world data-analysis tasks built from Kaggle notebooks spanning 2,225 interactive turns, finding that long-horizon errors account for 52-69% of agent failures and that maintaining a correct analytical state is the core bottleneck. | The work turns loop quality into a measurable task or score. Benchmark of 68 real-world data-analysis tasks built from Kaggle notebooks spanning 2,225 interactive turns, finding that long-horizon errors account for 52-69% of agent failures and that maintaining a correct analytical state is the core bottleneck. | Use LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,512 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1512 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2605.30434 | [2605.30434] LongDS-Bench: On the Failure of Long-Horizon Agentic Data Analysis | Real-world data analysis is inherently iterative, yet existing benchmarks mostly evaluate isolated or short interactive tasks, leaving agents' ability to track evolving analytical context over long horizons untested. We introduce LongDS, a benchmark for long-horizon, multi-turn data analysis where agents must maintain,... | Kewei Xu; Xiaoben Lu; Shuofei Qiao; Zihan Ding; Haoming Xu; Lei Liang; Ningyu Zhang | 2026-05-28 | 2026 | arXiv | arXiv | EMNLP 2026; Project Home: https://zjunlp.github.io/DataMind/ | cs.LG | arxiv-api | 2605.30434 | 2026-09-26T17:25:52 | ||||||||
ale-0806 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks | https://arxiv.org/abs/2602.16313 | external | arxiv.org | Multi-session benchmark of interdependent agentic tasks where agents must distill earlier sessions into memory and use it to guide later actions, showing that near-saturated scores on long-context memory benchmarks fail to transfer. | Multi-session benchmark of interdependent agentic tasks where agents must distill earlier sessions into memory and use it to guide later actions, showing that near-saturated scores on long-context memory benchmarks fail to transfer. | Multi-session benchmark of interdependent agentic tasks where agents must distill earlier sessions into memory and use it to guide later actions, showing that near-saturated scores on long-context memory benchmarks fail to transfer. | The work turns loop quality into a measurable task or score. Multi-session benchmark of interdependent agentic tasks where agents must distill earlier sessions into memory and use it to guide later actions, showing that near-saturated scores on long-context memory benchmarks fail to transfer. | Use MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,513 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1513 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2602.16313 | [2602.16313] MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks | Existing evaluations of agents with memory typically assess memorization and action in isolation. One class of benchmarks evaluates memorization by testing recall of past conversations or text but fails to capture how memory is used to guide future decisions. Another class focuses on agents acting in single-session tas... | Zexue He; Yu Wang; Churan Zhi; Yuanzhe Hu; Tzu-Ping Chen; Lang Yin; Ze Chen; Tong Arthur Wu; Siru Ouyang; Zihan Wang; Jiaxin Pei; Julian McAuley; Yejin Choi; Alex Pentland | 2026-02-18 | 2026 | arXiv | arXiv | ICML 2026 | cs.CL | arxiv-api | 2602.16313 | 2026-09-26T17:25:52 | ||||||||
ale-0807 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Momento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic Conversations | https://arxiv.org/abs/2606.00832 | external | arxiv.org | Benchmark for persistent, tool-mediated task completion across multiple sessions, finding that agents fail by treating prior-session history as current context instead of stale state that needs re-validation. | Benchmark for persistent, tool-mediated task completion across multiple sessions, finding that agents fail by treating prior-session history as current context instead of stale state that needs re-validation. | Benchmark for persistent, tool-mediated task completion across multiple sessions, finding that agents fail by treating prior-session history as current context instead of stale state that needs re-validation. | The work turns loop quality into a measurable task or score. Benchmark for persistent, tool-mediated task completion across multiple sessions, finding that agents fail by treating prior-session history as current context instead of stale state that needs re-validation. | Use Momento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic Conversations to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,514 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1514 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;context;verification;state;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2606.00832 | [2606.00832] Momento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic Conversations | Recent advances in agentic AI have enabled agents to complete complex tasks through tool use, reasoning, and multi-step planning. Yet existing benchmarks evaluate agents within a single session, ignoring past actions, stated preferences, and prior decisions that agents must integrate to fulfill personalized user goals.... | Adril Putra Merin; David Anugraha; Ayu Purwarianti; Genta Indra Winata | 2026-05-30 | 2026 | arXiv | arXiv | Preprint | cs.CL | arxiv-api | 2606.00832 | 2026-09-26T17:25:52 | ||||||||
ale-0808 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows | https://arxiv.org/abs/2605.14678 | external | arxiv.org | Benchmark of 100 multi-turn tasks across 5 user personas with hidden intents, inter-task dependencies, and cross-session continuity, measuring agent proactivity separately from task completion in long-horizon trajectories. | Benchmark of 100 multi-turn tasks across 5 user personas with hidden intents, inter-task dependencies, and cross-session continuity, measuring agent proactivity separately from task completion in long-horizon trajectories. | Benchmark of 100 multi-turn tasks across 5 user personas with hidden intents, inter-task dependencies, and cross-session continuity, measuring agent proactivity separately from task completion in long-horizon trajectories. | The work turns loop quality into a measurable task or score. Benchmark of 100 multi-turn tasks across 5 user personas with hidden intents, inter-task dependencies, and cross-session continuity, measuring agent proactivity separately from task completion in long-horizon trajectories. | Use π-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,515 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1515 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2605.14678 | [2605.14678] $π$-Bench: Evaluating Proactive Personal Assistant Agents in Long-Horizon Workflows | The rise of personal assistant agents, e.g., OpenClaw, highlights the growing potential of large language models to support users across everyday life and work. A core challenge in these settings is proactive assistance, since users often begin with underspecified requests and leave important needs, constraints, or pre... | Haoran Zhang; Luxin Xu; Zhilin Wang; Runquan Gui; Shunkai Zhang; Haodi Lei; Zihao He; Bingsu He; Chicheng Qin; Tong Zhu; Xiaoye Qu; Yang Yang; Yu Cheng; Yafu Li | 2026-05-14 | 2026 | arXiv | arXiv | 44 pages | cs.AI | arxiv-api | 2605.14678 | 2026-09-26T17:25:52 | ||||||||
ale-0809 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Can LLM Agents Be CFOs? Benchmarking Long-Horizon Resource Allocation | https://arxiv.org/abs/2603.23638 | external | arxiv.org | A 132-month CFO simulation where agents repeat a monthly cycle of liquidity management, financial closings, and financing decisions with compounding state, and only 15.4% of trials survive the full horizon. | A 132-month CFO simulation where agents repeat a monthly cycle of liquidity management, financial closings, and financing decisions with compounding state, and only 15.4% of trials survive the full horizon. | A 132-month CFO simulation where agents repeat a monthly cycle of liquidity management, financial closings, and financing decisions with compounding state, and only 15.4% of trials survive the full horizon. | The work targets tasks that exceed a single context window or prompt session. A 132-month CFO simulation where agents repeat a monthly cycle of liquidity management, financial closings, and financing decisions with compounding state, and only 15.4% of trials survive the full horizon. | Use Can LLM Agents Be CFOs? Benchmarking Long-Horizon Resource Allocation to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,516 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1516 | Verify | verify | Gate progress with tests, evals, and evidence. | state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2603.23638 | [2603.23638] Can LLM Agents Be CFOs? Benchmarking Long-Horizon Resource Allocation in an Uncertain Enterprise Environment | Large language model (LLM) agents are increasingly tested on complex tasks, but their ability to allocate scarce resources over long horizons remains unclear. Unlike reactive tasks with immediate feedback, this setting requires agents to make binding commitments under partial observability, delayed consequences, hard r... | Yi Han; Yan Wang; Lingfei Qian; Haohang Li; Yupeng Cao; Yueru He; Xueqing Peng; Nanhan Shen; Yitao Xu; Yankai Chen; Dongji Feng; Jimin Huang; Xue Liu; Jian-Yun Nie; Sophia Ananiadou | 2026-03-24 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2603.23638 | 2026-09-26T17:25:52 | |||||||||
ale-0810 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer | https://arxiv.org/abs/2607.05202 | external | arxiv.org | Benchmark isolating whether agents transfer reusable procedures such as searching, debugging, and verification across episodes in four long-horizon domains linked by ability graphs, finding curated experience transfers but no automatic method yields consistent gains. | Benchmark isolating whether agents transfer reusable procedures such as searching, debugging, and verification across episodes in four long-horizon domains linked by ability graphs, finding curated experience transfers but no automatic method yields consistent gains. | Benchmark isolating whether agents transfer reusable procedures such as searching, debugging, and verification across episodes in four long-horizon domains linked by ability graphs, finding curated experience transfers but no automatic method yields consistent gains. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Benchmark isolating whether agents transfer reusable procedures such as searching, debugging, and verification across episodes in four long-horizon domains linked by ability graphs, finding curated experience transfers but no automat... | Use EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,517 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1517 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.05202 | [2607.05202] EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer | Agent self-evolution in long-horizon LLM systems is largely procedural: useful experience is not merely stored information, but reusable procedures for searching, debugging, and verification. Yet current evaluations do not isolate this form of transfer. Agent benchmarks test single-episode task solving; memory benchmar... | Xingze Gao; Chuanrui Hu; Hongda Chen; Pengfei Yao; Zhao Wang; Yi Bai; Zhengwei Wu; Yunyun Han; Xiaofeng Cong; Jie Gui; Yafeng Deng; Teng Li | 2026-07-06 | 2026 | arXiv | arXiv | 15 pages, 2 figures, 8 tables | cs.AI | arxiv-api | 2607.05202 | 2026-09-26T17:25:52 | ||||||||
ale-0811 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents | https://arxiv.org/abs/2607.02255 | external | arxiv.org | Bounded-memory testbed built on Slay the Spire 2 where every agent decision is made from a fresh prompt assembled by typed retrieval over recorded state, keeping prompt size bounded across runs of any length, with 298 documented trajectories released. | Bounded-memory testbed built on Slay the Spire 2 where every agent decision is made from a fresh prompt assembled by typed retrieval over recorded state, keeping prompt size bounded across runs of any length, with 298 documented trajectories released. | Bounded-memory testbed built on Slay the Spire 2 where every agent decision is made from a fresh prompt assembled by typed retrieval over recorded state, keeping prompt size bounded across runs of any length, with 298 documented trajectories released. | Persistent memory is treated as an external runtime artifact. Bounded-memory testbed built on Slay the Spire 2 where every agent decision is made from a fresh prompt assembled by typed retrieval over recorded state, keeping prompt size bounded across runs of any length, with 298 documented trajectories released. | Use AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,518 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1518 | Verify | verify | Gate progress with tests, evals, and evidence. | context;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.02255 | [2607.02255] AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents | Memory for a long-horizon LLM agent is a contract about what each future decision is allowed to see. The simplest contract appends past observations, tool calls, and reflections to every prompt, which makes prior context easy to access but also turns it into a jumbled mixture in which the effect of any single memory co... | Xiangchen Cheng; Yunwei Jiang; Jianwen Sun; Zizhen Li; Chuanhao Li; Xiangcheng Cao; Yihao Liu; Fanrui Zhang; Li Jin; Kaipeng Zhang | 2026-07-02 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.02255 | 2026-09-26T17:25:52 | |||||||||
ale-0812 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Is Three the Magic Number? An Empirical Evaluation of LLM-Based Repair Loops | https://arxiv.org/abs/2607.05197 | external | arxiv.org | Empirical evaluation of iteration budgets for generate-validate-repair loops across code generation, test generation, and translation, finding the first three to four iterations capture most gains and that orchestration and feedback design matter more than the model. | Empirical evaluation of iteration budgets for generate-validate-repair loops across code generation, test generation, and translation, finding the first three to four iterations capture most gains and that orchestration and feedback design matter more than the model. | Empirical evaluation of iteration budgets for generate-validate-repair loops across code generation, test generation, and translation, finding the first three to four iterations capture most gains and that orchestration and feedback design matter more than the model. | Evaluation data is used as the feedback signal for improving loop behavior. Empirical evaluation of iteration budgets for generate-validate-repair loops across code generation, test generation, and translation, finding the first three to four iterations capture most gains and that orchestration and feedback design matt... | Use Is Three the Magic Number? An Empirical Evaluation of LLM-Based Repair Loops to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.05197; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,519 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1519 | Verify | verify | Gate progress with tests, evals, and evidence. | delegation;verification;budget | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.05197 | [2607.05197] Is Three the Magic Number? An Empirical Evaluation of LLM-Based Repair Loops | Iterative repair loops have become a core design pattern in LLM-based software engineering systems. These workflows repeatedly generate, validate, and repair artifacts using feedback such as compiler errors or test failures. Despite their widespread use, the impact of repair-loop iteration limits remains poorly underst... | Tobias Kiecker; Eik Reichmann; Hosung Kang; Gabin An; Lars Grunske | 2026-07-06 | 2026 | arXiv | arXiv | 4 Pages (+1 for references), NIER Paper | cs.SE | arxiv-api | 2607.05197 | 2026-09-26T17:25:52 | ||||||||
ale-0813 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks | https://arxiv.org/abs/2607.07946 | external | arxiv.org | 113 from-scratch, contamination-free long-horizon software-engineering tasks with custom verifiers that accept any correct implementation, built to sidestep SWE-bench-style pretraining recall and better differentiate frontier coding agents. | 113 from-scratch, contamination-free long-horizon software-engineering tasks with custom verifiers that accept any correct implementation, built to sidestep SWE-bench-style pretraining recall and better differentiate frontier coding agents. | 113 from-scratch, contamination-free long-horizon software-engineering tasks with custom verifiers that accept any correct implementation, built to sidestep SWE-bench-style pretraining recall and better differentiate frontier coding agents. | The work targets tasks that exceed a single context window or prompt session. 113 from-scratch, contamination-free long-horizon software-engineering tasks with custom verifiers that accept any correct implementation, built to sidestep SWE-bench-style pretraining recall and better differentiate frontier coding agents. | Use DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,520 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1520 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.07946 | [2607.07946] DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks | DeepSWE is a benchmark of 113 original, long-horizon software engineering tasks for evaluating coding agents. Most public agentic coding benchmarks follow SWE-bench in mining merged fixes from public GitHub repositories, which creates two problems: the fixes and their discussion were likely seen during pretraining, so ... | Wenqi Huang; Charley Lee; Leonard Tng; Serena Ge | 2026-07-08 | 2026 | arXiv | arXiv | 32 pages, 10 figures. Code and data: https://github.com/datacurve-ai/deep-swe ; https://deepswe.datacurve.ai/ | cs.SE | arxiv-api | 2607.07946 | 2026-09-26T17:25:52 | ||||||||
ale-0814 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | PERFOPT-Bench: Evaluating Coding Agents on Software Performance Optimization | https://arxiv.org/abs/2607.07744 | external | arxiv.org | Benchmarks the profile-diagnose-edit-verify loop where the verifier is a profiler rather than a test suite: agents must deliver measured, reproducible speedups without breaking correctness, and across seven agent configurations the framework choice shifts results even with identical models. | Benchmarks the profile-diagnose-edit-verify loop where the verifier is a profiler rather than a test suite: agents must deliver measured, reproducible speedups without breaking correctness, and across seven agent configurations the framework choice shifts results even with identical models. | Benchmarks the profile-diagnose-edit-verify loop where the verifier is a profiler rather than a test suite: agents must deliver measured, reproducible speedups without breaking correctness, and across seven agent configurations the framework choice shifts results even with identical models. | Verification is promoted from a final check to a loop-control signal. Benchmarks the profile-diagnose-edit-verify loop where the verifier is a profiler rather than a test suite: agents must deliver measured, reproducible speedups without breaking correctness, and across seven agent configurations the framework choice s... | Use PERFOPT-Bench: Evaluating Coding Agents on Software Performance Optimization to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,521 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1521 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.07744 | [2607.07744] PERFOPT-Bench: Evaluating Coding Agents on Software Performance Optimization | Coding-agent benchmarks have largely measured whether agents can produce functionally correct patches, but production software also demands measurable speedups on real execution targets. Performance optimization is a distinct agentic task: agents must profile executions, diagnose cross-layer bottlenecks, edit code with... | Yingyun Cui; Yi Xie; Piaohong Wang; Jiawei Ma; Bo Liu; Liangliang Cao | 2026-07-08 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2607.07744 | 2026-09-26T17:25:52 | |||||||||
ale-0815 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Benchmarking coding agents on Databricks' multi-million line codebase | https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase | external | www.databricks.com | Databricks engineering post (July 8, 2026, authors including Matei Zaharia and Patrick Wendell) on an internal benchmark built from real merged PRs with test-suite verification, finding that models cluster into three capability tiers, token price is a poor proxy for end-to-end task cost, and harness choice matters, wit... | Databricks engineering post (July 8, 2026, authors including Matei Zaharia and Patrick Wendell) on an internal benchmark built from real merged PRs with test-suite verification, finding that models cluster into three capability tiers, token price is a poor proxy for end-to-end task cost, and harness choice matters, wit... | Databricks engineering post (July 8, 2026, authors including Matei Zaharia and Patrick Wendell) on an internal benchmark built from real merged PRs with test-suite verification, finding that models cluster into three capability tiers, token price is a poor proxy for end-to-end task cost, and harness choice matters, wit... | Verification is promoted from a final check to a loop-control signal. Databricks engineering post (July 8, 2026, authors including Matei Zaharia and Patrick Wendell) on an internal benchmark built from real merged PRs with test-suite verification, finding that models cluster into three capability tiers, token price is ... | Use Benchmarking coding agents on Databricks' multi-million line codebase to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,522 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1522 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase | Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase | Databricks Blog | Databricks shares results from its internal coding benchmark, evaluating coding agents on a multi-million line codebase to optimize engineering cost and performance. | 2026 | 2026 | Databricks | html-meta | 2026-09-26T17:25:52 | |||||||||||||
ale-0816 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks | https://arxiv.org/abs/2607.08768 | external | arxiv.org | Capability-driven benchmark of 400 bilingual tasks for proactive agents operating everyday tools in live Docker environments with step-level checkpoints, decomposed into five foundational capabilities, skill usage, exploration, long-context reasoning, multimodal understanding, and cross-platform coordination, so failur... | Capability-driven benchmark of 400 bilingual tasks for proactive agents operating everyday tools in live Docker environments with step-level checkpoints, decomposed into five foundational capabilities, skill usage, exploration, long-context reasoning, multimodal understanding, and cross-platform coordination, so failur... | Capability-driven benchmark of 400 bilingual tasks for proactive agents operating everyday tools in live Docker environments with step-level checkpoints, decomposed into five foundational capabilities, skill usage, exploration, long-context reasoning, multimodal understanding, and cross-platform coordination, so failur... | Checkpointed state makes long-running agent work recoverable across failures. Capability-driven benchmark of 400 bilingual tasks for proactive agents operating everyday tools in live Docker environments with step-level checkpoints, decomposed into five foundational capabilities, skill usage, exploration, long-context r... | Use UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,523 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1523 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;context;verification;state;escalation | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.08768 | [2607.08768] UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks | The rapid development of large language models and multimodal large language models has accelerated the emergence of proactive agents capable of operating everyday tools and assisting users in real-world environments. However, existing benchmarks struggle to evaluate such agents effectively, as they often rely on sandb... | Zhekai Chen; Chengqi Duan; Kaiyue Sun; Bohao Li; Yuqing Wang; Manyuan Zhang; Xihui Liu | 2026-07-09 | 2026 | arXiv | arXiv | Project Page: https://uniclawbench.github.io | GitHub Repo: https://github.com/HKU-MMLab/UniClawBench | cs.CL | arxiv-api | 2607.08768 | 2026-09-26T17:25:52 | ||||||||
ale-0817 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | SLBench: Evaluating How LLM Agents Follow Logical Relations in Skills | https://arxiv.org/abs/2607.09016 | external | arxiv.org | Benchmark for whether agent loops respect the logical relations inside skill files (preconditions, constraints, fallbacks): 70% of 5,000+ public skills contain at least one such relation, and on 86 executable cases leading coding agents show unsafe-behavior rates up to 70%, with an inference-time scaffold cutting viola... | Benchmark for whether agent loops respect the logical relations inside skill files (preconditions, constraints, fallbacks): 70% of 5,000+ public skills contain at least one such relation, and on 86 executable cases leading coding agents show unsafe-behavior rates up to 70%, with an inference-time scaffold cutting viola... | Benchmark for whether agent loops respect the logical relations inside skill files (preconditions, constraints, fallbacks): 70% of 5,000+ public skills contain at least one such relation, and on 86 executable cases leading coding agents show unsafe-behavior rates up to 70%, with an inference-time scaffold cutting viola... | The work turns loop quality into a measurable task or score. Benchmark for whether agent loops respect the logical relations inside skill files (preconditions, constraints, fallbacks): 70% of 5,000+ public skills contain at least one such relation, and on 86 executable cases leading coding agents show unsafe-behavior r... | Use SLBench: Evaluating How LLM Agents Follow Logical Relations in Skills to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,524 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1524 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.09016 | [2607.09016] SLBench: Evaluating How LLM Agents Follow Logical Relations in Skills | Agent skills extend LLM agents with reusable procedures, tools, and domain-specific workflows, but their safety depends on resolving dependencies among interacting instructions. We introduce SkillLogic, a framework for analyzing logical relations in skill files and constructing executable tests from them. Our taxonomy ... | Xuan Chen; Chengpeng Wang; Lu Yan; Xiangyu Zhang | 2026-07-10 | 2026 | arXiv | arXiv | cs.CR | arxiv-api | 2607.09016 | 2026-09-26T17:25:52 | |||||||||
ale-0818 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution | https://arxiv.org/abs/2603.13428 | external | arxiv.org | Commit-history-derived milestone task streams where agents must preserve system integrity across successive runs - frontier-model scores collapse from >80% on isolated tasks to at most 38% in continuous settings, quantifying the error-accumulation gap loop engineering targets. | Commit-history-derived milestone task streams where agents must preserve system integrity across successive runs - frontier-model scores collapse from >80% on isolated tasks to at most 38% in continuous settings, quantifying the error-accumulation gap loop engineering targets. | Commit-history-derived milestone task streams where agents must preserve system integrity across successive runs - frontier-model scores collapse from >80% on isolated tasks to at most 38% in continuous settings, quantifying the error-accumulation gap loop engineering targets. | Links loop design to measurable tasks where progress and failure can be compared. Commit-history-derived milestone task streams where agents must preserve system integrity across successive runs - frontier-model scores collapse from >80% on isolated tasks to at most 38% in continuous settings, quantifying the error-acc... | Use SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,525 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1525 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2603.13428 | [2603.13428] SWE-Milestone: Evaluating AI Agents on Continuous Software Evolution | Real-world software must continuously evolve to meet ever-changing and open-ended requirements. AI agents, increasingly deployed as long-running systems, are now entrusted to drive this evolution. Yet, existing benchmarks evaluate agents on isolated, one-off coding tasks, neglecting the temporal dependencies and techni... | Gangda Deng; Zhaoling Chen; Zhongming Yu; Haoyang Fan; Yuhong Liu; Yuxin Yang; Dhruv Parikh; Rajgopal Kannan; Le Cong; Mengdi Wang; Qian Zhang; Viktor Prasanna; Xiangru Tang; Xingyao Wang | 2026 | 2026 | Proceedings of the 43rd International Conference on Machine Learning (ICML) | PMLR | Accepted at Proceedings of the 43rd International Conference on Machine Learning (ICML); the linked arXiv record is the available paper version. | cs.SE | Current arXiv acceptance note and official project record | 2603.13428 | 2026-09-26T17:25:52 | ||||||||
ale-0819 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AgentAbstain: Do LLM Agents Know When Not to Act? | https://arxiv.org/abs/2607.10059 | external | arxiv.org | Benchmark measuring whether agents correctly abstain from acting when a task is underspecified, unsafe, or impossible, rather than proceeding anyway, a capability every unattended loop depends on for safe exits. | Benchmark measuring whether agents correctly abstain from acting when a task is underspecified, unsafe, or impossible, rather than proceeding anyway, a capability every unattended loop depends on for safe exits. | Benchmark measuring whether agents correctly abstain from acting when a task is underspecified, unsafe, or impossible, rather than proceeding anyway, a capability every unattended loop depends on for safe exits. | The work turns loop quality into a measurable task or score. Benchmark measuring whether agents correctly abstain from acting when a task is underspecified, unsafe, or impossible, rather than proceeding anyway, a capability every unattended loop depends on for safe exits. | Use AgentAbstain: Do LLM Agents Know When Not to Act? to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,526 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1526 | 2026-07-15 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.10059 | [2607.10059] AgentAbstain: Do LLM Agents Know When Not to Act? | Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents know when to abstain. This gap poses real risks: under ambiguity, conflicting constraints, or tool failures, agents may execute unintended ... | Xun Liu; Yi Evie Zhang; Vira Kasprova; Parisa Rabbani; Pardis Sadat Zahraei; Tianyu Zhang; Ali Ebrahimpour-Boroojeny; Varun Chandrasekaran | 2026-07-11 | 2026 | arXiv | arXiv | 56 pages, 13 figures | cs.AI | arxiv-api | 2607.10059 | 2026-09-26T17:25:52 | |||||||
ale-0820 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy | https://arxiv.org/abs/2607.10526 | external | arxiv.org | Benchmark showing that once a stateful personal agent is nudged into a sycophantic stance, it persists across later sessions through memory, so single-turn sycophancy tests understate the risk in long-running agents. | Benchmark showing that once a stateful personal agent is nudged into a sycophantic stance, it persists across later sessions through memory, so single-turn sycophancy tests understate the risk in long-running agents. | Benchmark showing that once a stateful personal agent is nudged into a sycophantic stance, it persists across later sessions through memory, so single-turn sycophancy tests understate the risk in long-running agents. | The work turns loop quality into a measurable task or score. Benchmark showing that once a stateful personal agent is nudged into a sycophantic stance, it persists across later sessions through memory, so single-turn sycophancy tests understate the risk in long-running agents. | Use Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,527 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1527 | 2026-07-15 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.10526 | [2607.10526] Agents Don't Just Agree, They Remember: Benchmarking Persistent Sycophancy in Stateful Personal Agents | Stateful personal agents increasingly maintain long-term user profiles, episodic memories, and reusable skills. This persistence turns conversational sycophancy into a state-writing failure: accepted user-centric claims can be committed as lasting preferences, background facts, or workflows and later reused after the o... | Xutao Mao; Liangjie Zhao; Leyao Wang; Rui Qian; Qiang Huang; Wentao Wang; Bo Han; Xiang Zheng; Cong Wang | 2026-07-12 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.10526 | 2026-09-26T17:25:52 | ||||||||
ale-0821 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Set-shifting Behavioral Test for Harnessed Agents | https://arxiv.org/abs/2607.13396 | external | arxiv.org | Tests whether harnessed agents adapt when hidden tool reliability changes, paired with no-shift controls that separate genuine adaptation failures from ordinary task errors and expose routine lock-in. | Tests whether harnessed agents adapt when hidden tool reliability changes, paired with no-shift controls that separate genuine adaptation failures from ordinary task errors and expose routine lock-in. | Tests whether harnessed agents adapt when hidden tool reliability changes, paired with no-shift controls that separate genuine adaptation failures from ordinary task errors and expose routine lock-in. | Links loop design to measurable tasks where progress and failure can be compared. Tests whether harnessed agents adapt when hidden tool reliability changes, paired with no-shift controls that separate genuine adaptation failures from ordinary task errors and expose routine lock-in. | Use Set-shifting Behavioral Test for Harnessed Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,528 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1528 | 2026-07-17 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.13396 | [2607.13396] Set-shifting Behavioral Test for Harnessed Agents | What happens to an LLM agent's tool choice when the reliable tool silently changes within an ongoing session? We borrow the notion of set-shifting from cognitive psychology to study how well agents adapt to hidden reliability shifts. Our cognitive test for LLM agents mounts libraries of redundant tools and skills, in w... | Ye Ziwei | 2026-07-15 | 2026 | arXiv | arXiv | Accepted at COLM 2026 Workshop on Agent Behavior | cs.AI | arxiv-api | 2607.13396 | 2026-09-26T17:25:52 | |||||||
ale-0822 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | MCPEvol-Bench: Benchmarking LLM Agent Performance Across Dynamic Evolutions of MCP Servers | https://arxiv.org/abs/2607.14642 | external | arxiv.org | Applies 11 server-evolution mutations to 123 MCP servers and evaluates 12 LLMs; reported performance drops include 13.7% for GPT-5.4 and 14.4% for Claude Sonnet 4.6, quantifying tool-interface drift. | Applies 11 server-evolution mutations to 123 MCP servers and evaluates 12 LLMs; reported performance drops include 13.7% for GPT-5.4 and 14.4% for Claude Sonnet 4.6, quantifying tool-interface drift. | Applies 11 server-evolution mutations to 123 MCP servers and evaluates 12 LLMs; reported performance drops include 13.7% for GPT-5.4 and 14.4% for Claude Sonnet 4.6, quantifying tool-interface drift. | Links loop design to measurable tasks where progress and failure can be compared. Applies 11 server-evolution mutations to 123 MCP servers and evaluates 12 LLMs; reported performance drops include 13.7% for GPT-5.4 and 14.4% for Claude Sonnet 4.6, quantifying tool-interface drift. | Use MCPEvol-Bench: Benchmarking LLM Agent Performance Across Dynamic Evolutions of MCP Servers to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,529 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1529 | 2026-07-17 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.14642 | [2607.14642] MCPEvol-Bench: Benchmarking LLM Agent Performance Across Dynamic Evolutions of MCP Servers | As Model Context Protocol (MCP) servers emerge as the core infrastructure for connecting LLMs with external tools, existing benchmarks leverage real-world MCP servers to evaluate LLM agents' tool-using capabilities. However, these benchmarks overlook the continuous evolution of tool interfaces and functionalities withi... | Huanxi Liu; Kun Hu; Jiaqi Liao; Qiang Wang; Pengfei Qian; YuanZhao Zhai; Dawei Feng; Bo Ding; Huaimin Wang | 2026-07-16 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.14642 | 2026-09-26T17:25:52 | ||||||||
ale-0823 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization | https://arxiv.org/abs/2607.15205 | external | arxiv.org | Provides 652 issue-PR instances across 23 languages, seven image categories, and four relevance levels; the strongest evaluated agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, leaving substantial room for visual-evidence-aware repair loops. | Provides 652 issue-PR instances across 23 languages, seven image categories, and four relevance levels; the strongest evaluated agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, leaving substantial room for visual-evidence-aware repair loops. | Provides 652 issue-PR instances across 23 languages, seven image categories, and four relevance levels; the strongest evaluated agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, leaving substantial room for visual-evidence-aware repair loops. | The work turns loop quality into a measurable task or score. Provides 652 issue-PR instances across 23 languages, seven image categories, and four relevance levels; the strongest evaluated agent reaches 38.96 file Acc@5 and 22.45 function Acc@10, leaving substantial room for visual-evidence-aware repair loops. | Use MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,530 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1530 | 2026-07-17 | Verify | verify | Gate progress with tests, evals, and evidence. | intake;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.15205 | [2607.15205] MM-IssueLoc: A Controlled Benchmark for Evaluating Visual Evidence in Multimodal Repository-Level Issue Localization | Real repository issues routinely include visual evidence such as screenshots, error dialogs, rendered UI states, and logs, yet repository-level issue localization is evaluated mostly as a text-only task. Existing multimodal SE benchmarks evaluate end-to-end repair, entangling localization with patch synthesis and obscu... | Shaoxiong Zhan; Shi Hu; Boyu Feng; Hai Lin; Andrew Gong; Zhengda Zhou; Jiaying Zhou; Yunyun Hou; Hao Su; Hai-Tao Zheng | 2026-07-16 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2607.15205 | 2026-09-26T17:25:52 | ||||||||
ale-0824 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | ToolVerse: Unlocking Massive Environments and Long-Horizon Tasks for Agentic Reinforcement Learning | https://arxiv.org/abs/2607.15660 | external | arxiv.org | Builds executable long-horizon training environments from nearly 400 MCP servers and about 4,500 tools, generates tasks from tool-dependency graphs, and introduces turn-aware credit assignment for agentic reinforcement learning in large, dynamic tool spaces. | Builds executable long-horizon training environments from nearly 400 MCP servers and about 4,500 tools, generates tasks from tool-dependency graphs, and introduces turn-aware credit assignment for agentic reinforcement learning in large, dynamic tool spaces. | Builds executable long-horizon training environments from nearly 400 MCP servers and about 4,500 tools, generates tasks from tool-dependency graphs, and introduces turn-aware credit assignment for agentic reinforcement learning in large, dynamic tool spaces. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Builds executable long-horizon training environments from nearly 400 MCP servers and about 4,500 tools, generates tasks from tool-dependency graphs, and introduces turn-aware credit assignment for agentic reinforcement learning in la... | Use ToolVerse: Unlocking Massive Environments and Long-Horizon Tasks for Agentic Reinforcement Learning to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,531 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1531 | 2026-07-20 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.15660 | [2607.15660] ToolVerse: Unlocking Massive Environments and Long-Horizon Tasks for Agentic Reinforcement Learning | While LLM agents demonstrate strong reasoning abilities in compact and well-defined scenarios, they struggle to maintain robustness and effectiveness when faced with large-scale, diverse, and dynamic real-world environments that demand seamless tool integration. To address this gap, we introduce ToolVerse, a comprehens... | Shuaiyu Zhou; Fengpeng Yue; Zengjie Hu; Yuanzhe Shen; Chenyang Zhang; feng hong; Cao Liu; Ke Zeng | 2026-07-17 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.15660 | 2026-09-26T17:25:52 | ||||||||
ale-0825 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interaction and MCP-Augmented Tasks | https://aclanthology.org/2026.acl-long.278/ | external | aclanthology.org | Evaluates long-horizon mobile work that requires clarification and MCP tool use in reproducible self-hosted environments with deterministic backend, storage, and callback verification. | Evaluates long-horizon mobile work that requires clarification and MCP tool use in reproducible self-hosted environments with deterministic backend, storage, and callback verification. | Evaluates long-horizon mobile work that requires clarification and MCP tool use in reproducible self-hosted environments with deterministic backend, storage, and callback verification. | Verification is promoted from a final check to a loop-control signal. Evaluates long-horizon mobile work that requires clarification and MCP tool use in reproducible self-hosted environments with deterministic backend, storage, and callback verification. | Use MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interaction and MCP-Augmented Tasks to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,535 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1535 | 2026-07-18 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://aclanthology.org/2026.acl-long.278/ | MobileWorld: Benchmarking Autonomous Mobile Agents in Agent-User Interactive and MCP-Augmented Environments - ACL Anthology | Quyu Kong, Xu Zhang, Zhenyu Yang, Nolan Gao, Chen Liu, Panrong Tong, Chenglin Cai, Hanzhang Zhou, Jianan Zhang, Liangyu Chen, Zhidan Liu, Steven Hoi, Yue Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026. | Quyu Kong; Xu Zhang; Zhenyu Yang; Nolan Gao; Chen Liu; Panrong Tong; Chenglin Cai; Hanzhang Zhou; Jianan Zhang; Liangyu Chen; Zhidan Liu; Steven Hoi; Yue Wang | 2026 | 2026 | Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | ACL Anthology | 10.18653/v1/2026.acl-long.278 | html-meta | 2026-09-26T17:25:52 | |||||||||
ale-0826 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AGENCYBENCH: Benchmarking the Frontiers of Autonomous Agents | https://aclanthology.org/2026.acl-long.337/ | external | aclanthology.org | Provides 138 high-fidelity tasks across six capabilities whose average rollout exceeds one million tokens and 90 tool calls, exposing context-retention and long-horizon execution limits. | Provides 138 high-fidelity tasks across six capabilities whose average rollout exceeds one million tokens and 90 tool calls, exposing context-retention and long-horizon execution limits. | Provides 138 high-fidelity tasks across six capabilities whose average rollout exceeds one million tokens and 90 tool calls, exposing context-retention and long-horizon execution limits. | Context is managed as durable loop state rather than a single prompt payload. Provides 138 high-fidelity tasks across six capabilities whose average rollout exceeds one million tokens and 90 tool calls, exposing context-retention and long-horizon execution limits. | Use AGENCYBENCH: Benchmarking the Frontiers of Autonomous Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,536 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1536 | 2026-07-18 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;context;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://aclanthology.org/2026.acl-long.337/ | AgencyBench: Benchmarking the Frontiers of Autonomous Agents in 1M-Token Real-World Contexts - ACL Anthology | Keyu Li, Junhao Shi, Yang Xiao, Mohan Jiang, Jie Sun, Yunze Wu, Dayuan Fu, Shijie Xia, Xiaojie Cai, Tianze Xu, Weiye Si, Wenjie Li, Dequan Wang, Pengfei Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026. | Keyu Li; Junhao Shi; Yang Xiao; Mohan Jiang; Jie Sun; Yunze Wu; Dayuan Fu; Shijie Xia; Xiaojie Cai; Tianze Xu; Weiye Si; Wenjie Li; Dequan Wang; Pengfei Liu | 2026 | 2026 | Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | ACL Anthology | 10.18653/v1/2026.acl-long.337 | html-meta | 2026-09-26T17:25:52 | |||||||||
ale-0827 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading | https://arxiv.org/abs/2607.08964 | external | arxiv.org | 46 long-horizon terminal tasks across nine categories (experiment reproduction, software engineering, scientific computing, and more) graded with dense intermediate rewards and partial credit instead of binary pass/fail; runs average 231 episodes and 85 minutes, and the strongest of 15 frontier models reaches only 15.2... | 46 long-horizon terminal tasks across nine categories (experiment reproduction, software engineering, scientific computing, and more) graded with dense intermediate rewards and partial credit instead of binary pass/fail; runs average 231 episodes and 85 minutes, and the strongest of 15 frontier models reaches only 15.2... | 46 long-horizon terminal tasks across nine categories (experiment reproduction, software engineering, scientific computing, and more) graded with dense intermediate rewards and partial credit instead of binary pass/fail; runs average 231 episodes and 85 minutes, and the strongest of 15 frontier models reaches only 15.2... | The work targets tasks that exceed a single context window or prompt session. 46 long-horizon terminal tasks across nine categories (experiment reproduction, software engineering, scientific computing, and more) graded with dense intermediate rewards and partial credit instead of binary pass/fail; runs average 231 epis... | Use Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,537 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1537 | 2026-07-22 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.08964 | [2607.08964] Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading | AI agents have become capable of autonomously completing short, well-specified tasks. However, existing terminal benchmarks largely focus on simple problems that finish within minutes and are evaluated only by their final outcome. This setup overlooks intermediate progress and partial solutions, yielding sparse reward ... | Zongxia Li; Zhongzhi Li; Yucheng Shi; Ruhan Wang; Junyao Yang; Zhichao Liu; Xiyang Wu; Anhao Li; Yue Yu; Ninghao Liu; Lichao Sun; Haotao Mi; Leowei Liang | 2026-07-09 | 2026 | arXiv | arXiv | 17 pages | cs.AI | arxiv-api | 2607.08964 | 2026-09-26T17:25:52 | |||||||
ale-0828 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | PolyWorkBench: Benchmarking Multilingual Long-Horizon LLM Agents | https://arxiv.org/abs/2607.06008 | external | arxiv.org | 67 long-horizon workplace-workflow tasks across commerce, knowledge work, legal, localization, and manufacturing, graded by a hybrid of structural checks, executable verification, and LLM judging. Shows state-of-the-art agents degrade sharply when workflows mix languages versus monolingual runs, a multilingual coverage... | 67 long-horizon workplace-workflow tasks across commerce, knowledge work, legal, localization, and manufacturing, graded by a hybrid of structural checks, executable verification, and LLM judging. Shows state-of-the-art agents degrade sharply when workflows mix languages versus monolingual runs, a multilingual coverage... | 67 long-horizon workplace-workflow tasks across commerce, knowledge work, legal, localization, and manufacturing, graded by a hybrid of structural checks, executable verification, and LLM judging. Shows state-of-the-art agents degrade sharply when workflows mix languages versus monolingual runs, a multilingual coverage... | Verification is promoted from a final check to a loop-control signal. 67 long-horizon workplace-workflow tasks across commerce, knowledge work, legal, localization, and manufacturing, graded by a hybrid of structural checks, executable verification, and LLM judging. Shows state-of-the-art agents degrade sharply when wo... | Use PolyWorkBench: Benchmarking Multilingual Long-Horizon LLM Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,538 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1538 | 2026-07-22 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.06008 | [2607.06008] PolyWorkBench: Benchmarking LLM Agents for Cross-Lingual Long-Horizon Workflows | While Large Language Model (LLM) agents excel at monolingual long-horizon planning and tool use, enterprise workflows inherently require processing multilingual resources across extended trajectories. The interaction between multilinguality and long-horizon execution, however, remains underexplored. We introduce PolyWo... | Hongliang Li; Yijin Liu; Zhiwei Zhang; Zihe Liu; Xinyue Lou; Jinan Xu; Fandong Meng; Kaiyu Huang | 2026-07-07 | 2026 | arXiv | arXiv | 17 Pages, 5 figures | cs.AI | arxiv-api | 2607.06008 | 2026-09-26T17:25:52 | |||||||
ale-0829 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations | https://arxiv.org/abs/2607.12893 | external | arxiv.org | Evaluates agent memory by the operations that maintain it, remembering, forgetting, updating, reflecting, with structured operation traces rather than final-answer accuracy, exposing lifecycle failure modes that answer-level evals miss. Finds current memory systems remain far from uniformly reliable. | Evaluates agent memory by the operations that maintain it, remembering, forgetting, updating, reflecting, with structured operation traces rather than final-answer accuracy, exposing lifecycle failure modes that answer-level evals miss. Finds current memory systems remain far from uniformly reliable. | Evaluates agent memory by the operations that maintain it, remembering, forgetting, updating, reflecting, with structured operation traces rather than final-answer accuracy, exposing lifecycle failure modes that answer-level evals miss. Finds current memory systems remain far from uniformly reliable. | Evaluation data is used as the feedback signal for improving loop behavior. Evaluates agent memory by the operations that maintain it, remembering, forgetting, updating, reflecting, with structured operation traces rather than final-answer accuracy, exposing lifecycle failure modes that answer-level evals miss. Finds c... | Use MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,539 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1539 | 2026-07-22 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.12893 | [2607.12893] MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations | Long-term memory has become a foundational capability for LLM-based agents that accompany users across extended, multi-session interactions. Existing benchmarks, however, evaluate such memory almost exclusively through downstream question answering, scoring only the correctness of a final answer. This black-box formula... | Xixuan Hao; Zeyu Zhang; Zehao Lin; Yihang Sun; Ziliang Guo; Xichong Zhang; Yuxuan Liang; Feiyu Xiong; Zhiyu Li | 2026-07-14 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.12893 | 2026-09-26T17:25:52 | ||||||||
ale-0830 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities | https://arxiv.org/abs/2607.13705 | external | arxiv.org | Open-source evaluation infrastructure that factors agent evaluation into three independent components, Benchmark, Harness, and Environment, with a fault-tolerant asynchronous runtime and failure-mode diagnostics across 20+ benchmarks, making the harness a swappable, measurable variable rather than a confound. | Open-source evaluation infrastructure that factors agent evaluation into three independent components, Benchmark, Harness, and Environment, with a fault-tolerant asynchronous runtime and failure-mode diagnostics across 20+ benchmarks, making the harness a swappable, measurable variable rather than a confound. | Open-source evaluation infrastructure that factors agent evaluation into three independent components, Benchmark, Harness, and Environment, with a fault-tolerant asynchronous runtime and failure-mode diagnostics across 20+ benchmarks, making the harness a swappable, measurable variable rather than a confound. | Evaluation data is used as the feedback signal for improving loop behavior. Open-source evaluation infrastructure that factors agent evaluation into three independent components, Benchmark, Harness, and Environment, with a fault-tolerant asynchronous runtime and failure-mode diagnostics across 20+ benchmarks, making th... | Use AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.13705; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,540 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1540 | 2026-07-22 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.13705 | [2607.13705] AgentCompass: A Unified Evaluation Infrastructure for Agent Capabilities | As Large Language Models (LLMs) evolve into autonomous agents, the need for unified evaluation infrastructure becomes critical. However, current evaluation pipelines remain highly fragmented and tightly coupled, hindering reproducibility and causing redundant engineering. To address this, we introduce AgentCompass, an ... | Kai Chen; Zichen Ding; Jiaye Ge; Shufan Jiang; Mo Li; Qingqiu Li; Zehao Li; Zonglin Li; Tianhao Liang; Shudong Liu; Zerun Ma; Zixin Shang; Wenhui Tian; Zun Wang; Liwei Wu; Zhenyu Wu; Jun Xu; Bowen Yang; Dingbo Yuan; Qi Zhang; Songyang Zhang; Peiheng Zhou; Dongsheng Zhu | 2026-07-15 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.13705 | 2026-09-26T17:25:52 | ||||||||
ale-0831 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Baselines Before Architecture: Evaluating Coding Agents for Autonomous Penetration Testing | https://arxiv.org/abs/2607.13085 | external | arxiv.org | Controlled study on the 104-task XBOW benchmark showing plain coding agents already solve a large share of tasks attributed to specialized security harnesses, a methodological warning to establish model-matched baselines before crediting the harness/loop architecture for performance gains. | Controlled study on the 104-task XBOW benchmark showing plain coding agents already solve a large share of tasks attributed to specialized security harnesses, a methodological warning to establish model-matched baselines before crediting the harness/loop architecture for performance gains. | Controlled study on the 104-task XBOW benchmark showing plain coding agents already solve a large share of tasks attributed to specialized security harnesses, a methodological warning to establish model-matched baselines before crediting the harness/loop architecture for performance gains. | The work turns loop quality into a measurable task or score. Controlled study on the 104-task XBOW benchmark showing plain coding agents already solve a large share of tasks attributed to specialized security harnesses, a methodological warning to establish model-matched baselines before crediting the harness/loop arch... | Use Baselines Before Architecture: Evaluating Coding Agents for Autonomous Penetration Testing to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.13085; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,541 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1541 | 2026-07-22 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.13085 | [2607.13085] Baselines Before Architecture: Evaluating Coding Agents for Autonomous Penetration Testing | Recent autonomous penetration testing papers report high benchmark scores while adding multi-component security harnesses around frontier LLMs. Because these systems often change both architecture and backbone model, it is difficult to tell how much performance comes from the harness rather than from the underlying mod... | Ananda Dhakal; Krish Neupane; Aarjan Chaudhary | 2026-07-13 | 2026 | arXiv | arXiv | cs.CR | arxiv-api | 2607.13085 | 2026-09-26T17:25:52 | ||||||||
ale-0832 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction | https://arxiv.org/abs/2607.20911 | external | arxiv.org | Tencent's coding-agent evaluation suite spanning Code, Web, Office, and Security domains, where every task is reverse-engineered from a real commit, PR, or business scenario and rewritten as a short colloquial role-played request, contamination resistance comes from this construction plus dataset versioning rather than... | Tencent's coding-agent evaluation suite spanning Code, Web, Office, and Security domains, where every task is reverse-engineered from a real commit, PR, or business scenario and rewritten as a short colloquial role-played request, contamination resistance comes from this construction plus dataset versioning rather than... | Tencent's coding-agent evaluation suite spanning Code, Web, Office, and Security domains, where every task is reverse-engineered from a real commit, PR, or business scenario and rewritten as a short colloquial role-played request, contamination resistance comes from this construction plus dataset versioning rather than... | Packages the evidence as queryable CSV and JSONL rather than only a rendered page. Tencent's coding-agent evaluation suite spanning Code, Web, Office, and Security domains, where every task is reverse-engineered from a real commit, PR, or business scenario and rewritten as a short colloquial role-played request, contam... | Use Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,542 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1542 | 2026-07-24 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.20911 | [2607.20911] Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction | We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for constructing and running distribution-informed coding-agent tasks across four wo... | Tencent WorkBuddy Bench Team; Siqi Cai; Shaopeng Chen; Xiang Fei; Yong Mao; Zihan Xu; Zhiheng Lyu; Zhijian Shao; Yuchen Shi; Shuwen Zhang; Chaofan Qiu; Linjie Che; Xiaoxi Zhao; Feng Wu; Kai Zhang; Chaofan Zhu; Yubin Qi; Xiaoyun Liang; Peijie Dong; Yunhao Zhang; Yuanjie Zhu; Ling Jiang; Xianjun Zhang; Zhehang Chu; Anyua... | 2026-07-23 | 2026 | arXiv | arXiv | 30 pages, 9 figures. Project page: https://workbuddybench.com/ ; code: https://github.com/Tencent/workbuddy-bench ; dataset: https://huggingface.co/datasets/tencent/workbuddy-bench | cs.CL | arxiv-api | 2607.20911 | 2026-09-26T17:25:52 | |||||||
ale-0833 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders | https://arxiv.org/abs/2607.21217 | external | arxiv.org | Benchmark that evaluates coding agents as interactive project builders: agents must turn ambiguous product intent into working software through planning, requirement clarification, tool use, debugging, and repository-level construction across multi-turn sessions with simulated users, graded on functional correctness, s... | Benchmark that evaluates coding agents as interactive project builders: agents must turn ambiguous product intent into working software through planning, requirement clarification, tool use, debugging, and repository-level construction across multi-turn sessions with simulated users, graded on functional correctness, s... | Benchmark that evaluates coding agents as interactive project builders: agents must turn ambiguous product intent into working software through planning, requirement clarification, tool use, debugging, and repository-level construction across multi-turn sessions with simulated users, graded on functional correctness, s... | The work turns loop quality into a measurable task or score. Benchmark that evaluates coding agents as interactive project builders: agents must turn ambiguous product intent into working software through planning, requirement clarification, tool use, debugging, and repository-level construction across multi-turn sessi... | Use ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,543 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1543 | 2026-07-24 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.21217 | [2607.21217] ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders | The recent emergence of vibe-coding workflows is changing what coding agents are expected to do. Instead of merely completing code under fully specified instructions, agents are increasingly expected to transform incomplete product intent into working software by combining various abilities including planning, requirem... | Zhongyuan Peng; Dan Huang; Chuyu Zhang; Caijun Xu; Changyi Xiao; Shibo Hong; David Lo; Lin Qiu; Xuezhi Cao; Jiyuan He; Yixin Cao | 2026-07-23 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.21217 | 2026-09-26T17:25:52 | ||||||||
ale-0834 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | GuardianAgentBench: Where Agents Fail and How to Guard Them | https://arxiv.org/abs/2607.20982 | external | arxiv.org | 580-scenario safety benchmark spanning six domains and five adversarial attack modes, run on LangChain, LlamaIndex, and Vectara agents; even the strongest configuration reaches only 74.8% accuracy, and a guardrail defense recovers 19.9% of failures at a 0.5% false-positive rate, mapping distinct failure regimes for gua... | 580-scenario safety benchmark spanning six domains and five adversarial attack modes, run on LangChain, LlamaIndex, and Vectara agents; even the strongest configuration reaches only 74.8% accuracy, and a guardrail defense recovers 19.9% of failures at a 0.5% false-positive rate, mapping distinct failure regimes for gua... | 580-scenario safety benchmark spanning six domains and five adversarial attack modes, run on LangChain, LlamaIndex, and Vectara agents; even the strongest configuration reaches only 74.8% accuracy, and a guardrail defense recovers 19.9% of failures at a 0.5% false-positive rate, mapping distinct failure regimes for gua... | The work turns loop quality into a measurable task or score. 580-scenario safety benchmark spanning six domains and five adversarial attack modes, run on LangChain, LlamaIndex, and Vectara agents; even the strongest configuration reaches only 74.8% accuracy, and a guardrail defense recovers 19.9% of failures at a 0.5% ... | Use GuardianAgentBench: Where Agents Fail and How to Guard Them to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,544 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1544 | 2026-07-24 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.20982 | [2607.20982] GuardianAgentBench: Where Agents Fail and How to Guard Them | As large language model agents increasingly operate autonomously with access to tools and external environments, ensuring their safe and reliable behavior becomes critical. We present GuardianAgentBench (GABench), a benchmark of 580 scenarios across six domains evaluated on three production-ready frameworks: LangChain,... | Vishal Ishwar Naik; Chenyu Xu; Donna Dong; Hussein Hassan; Abhishek Pradhan; Ofer Mendelevitch; Tallat Shafat; Humayun Irshad | 2026-07-23 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.20982 | 2026-09-26T17:25:52 | ||||||||
ale-0835 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | LLMs Get Lost in Evolving User Intent | https://arxiv.org/abs/2607.20734 | external | arxiv.org | Microsoft framework converting static single-turn tasks into dynamic multi-turn conversations where users disclose, revise, and reshape intent across the loop rather than specify it upfront, showing substantial performance drops across model families, as the interaction-loop counterpart to long-horizon execution and a ... | Microsoft framework converting static single-turn tasks into dynamic multi-turn conversations where users disclose, revise, and reshape intent across the loop rather than specify it upfront, showing substantial performance drops across model families, as the interaction-loop counterpart to long-horizon execution and a ... | Microsoft framework converting static single-turn tasks into dynamic multi-turn conversations where users disclose, revise, and reshape intent across the loop rather than specify it upfront, showing substantial performance drops across model families, as the interaction-loop counterpart to long-horizon execution and a ... | The work targets tasks that exceed a single context window or prompt session. Microsoft framework converting static single-turn tasks into dynamic multi-turn conversations where users disclose, revise, and reshape intent across the loop rather than specify it upfront, showing substantial performance drops across model ... | Use LLMs Get Lost in Evolving User Intent to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.20734; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,545 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1545 | 2026-07-24 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.20734 | [2607.20734] LLMs Get Lost in Evolving User Intent | As LLMs become more capable, they are increasingly deployed as collaborative agents, taking on user-delegated tasks through iterative interaction. Yet genuine interaction is inherently dynamic: users rarely specify their intent upfront, instead disclosing, revising, and reshaping it as the conversation unfolds. Despite... | Jihoon Tack; Philippe Laban; Jennifer Neville | 2026-07-22 | 2026 | arXiv | arXiv | 20 pages, 10 figures | cs.LG | arxiv-api | 2607.20734 | 2026-09-26T17:25:52 | |||||||
ale-0836 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management | https://arxiv.org/abs/2607.20764 | external | arxiv.org | Benchmark generator producing arbitrarily scalable task graphs, natural-language problems paired with executable Python solvers linked by typed intermediate values, with controllable length, dependencies, distractors, and exact automatic verification. | Benchmark generator producing arbitrarily scalable task graphs, natural-language problems paired with executable Python solvers linked by typed intermediate values, with controllable length, dependencies, distractors, and exact automatic verification. | Benchmark generator producing arbitrarily scalable task graphs, natural-language problems paired with executable Python solvers linked by typed intermediate values, with controllable length, dependencies, distractors, and exact automatic verification. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Benchmark generator producing arbitrarily scalable task graphs, natural-language problems paired with executable Python solvers linked by typed intermediate values, with controllable length, dependencies, distractors, and exact autom... | Use ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,546 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1546 | 2026-07-25 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.20764 | [2607.20764] ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management | We introduce ARBIGRAPH, a benchmark generator for evaluating whether tool-assisted language agents can retain, update, compose, and discard task-relevant context across extended reasoning workflows. ARBIGRAPH represents each task as a natural-language problem with an executable Python solver, and composes tasks through... | Pavel Golikov; Evgenii Opryshko; Gennady Pekhimenko; Mark C. Jeffrey | 2026-07-22 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.20764 | 2026-09-26T17:25:52 | ||||||||
ale-0837 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Code Monitor Red Teaming for Public-Test-Passing Code | https://arxiv.org/abs/2607.20852 | external | arxiv.org | Red-teams the monitor layer of coding-agent loops: can a weaker LLM verifier catch hidden bugs in code that already passes public tests? | Red-teams the monitor layer of coding-agent loops: can a weaker LLM verifier catch hidden bugs in code that already passes public tests? | Red-teams the monitor layer of coding-agent loops: can a weaker LLM verifier catch hidden bugs in code that already passes public tests?. | Verification is promoted from a final check to a loop-control signal. Red-teams the monitor layer of coding-agent loops: can a weaker LLM verifier catch hidden bugs in code that already passes public tests?. | Use Code Monitor Red Teaming for Public-Test-Passing Code to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.20852; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,547 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1547 | 2026-07-25 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.20852 | [2607.20852] Code Monitor Red Teaming for Public-Test-Passing Code | Visible tests are a common gate for LLM-generated code, but passing them does not certify specification correctness. We study a deployment-like monitoring problem: after code has passed public tests, can a weaker LLM verifier identify the residual hidden bugs? We introduce Code Monitor Red Teaming, a monitor-red-teamin... | Junchi Liao; Jiawen Deng; Fuji Ren | 2026-07-23 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.20852 | 2026-09-26T17:25:52 | ||||||||
ale-0838 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | SciExplore: Evaluating Autonomous Agents from Scientific Navigation to Information Integration | https://arxiv.org/abs/2607.20926 | external | arxiv.org | 103 expert-curated tasks across 10+ scientific disciplines evaluating long-horizon agent information-seeking loops, scientific database navigation, ambiguous literature retrieval, missing-reference completion, and cross-source knowledge synthesis. Evaluation of 10+ SOTA models shows accuracy collapsing as loop depth gr... | 103 expert-curated tasks across 10+ scientific disciplines evaluating long-horizon agent information-seeking loops, scientific database navigation, ambiguous literature retrieval, missing-reference completion, and cross-source knowledge synthesis. Evaluation of 10+ SOTA models shows accuracy collapsing as loop depth gr... | 103 expert-curated tasks across 10+ scientific disciplines evaluating long-horizon agent information-seeking loops, scientific database navigation, ambiguous literature retrieval, missing-reference completion, and cross-source knowledge synthesis. Evaluation of 10+ SOTA models shows accuracy collapsing as loop depth gr... | Evaluation data is used as the feedback signal for improving loop behavior. 103 expert-curated tasks across 10+ scientific disciplines evaluating long-horizon agent information-seeking loops, scientific database navigation, ambiguous literature retrieval, missing-reference completion, and cross-source knowledge synthes... | Use SciExplore: Evaluating Autonomous Agents from Scientific Navigation to Information Integration to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,548 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1548 | 2026-07-25 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://doi.org/10.18653/v1/2026.findings-acl.1117 | [2607.20926] SciExplore: Evaluating Autonomous Agents from Scientific Navigation to Information Integration | Scientific research involves complex information-seeking and reasoning workflows across heterogeneous sources. However, existing benchmarks primarily emphasize general-domain retrieval or static scientific question answering, and therefore fail to assess key capabilities required in realistic scientific research workfl... | Yinhao Tang; Youqing Fang; Yanan Sun; Wenran Liu; Weiming Zhang; Bin Liu; Kuikun Liu; Wenwei Zhang; Kai Chen | 2026 | 2026 | Findings of the Association for Computational Linguistics: ACL 2026 | Association for Computational Linguistics | 10.18653/v1/2026.findings-acl.1117 | Published in Findings of the Association for Computational Linguistics: ACL 2026; the linked arXiv record remains available for open access. | cs.AI | Crossref API + DOI record | 2607.20926 | 2026-09-26T17:25:52 | ||||||
ale-0839 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | RUMBA: Russian User Memory Benchmark | https://arxiv.org/abs/2607.21447 | external | arxiv.org | Long-term conversational memory benchmark built from timestamped multi-session user-assistant dialogues, with a fine-grained taxonomy of memory-centric questions spanning semantic type, session scope, temporal reasoning, and explicitness of temporal expressions. | Long-term conversational memory benchmark built from timestamped multi-session user-assistant dialogues, with a fine-grained taxonomy of memory-centric questions spanning semantic type, session scope, temporal reasoning, and explicitness of temporal expressions. | Long-term conversational memory benchmark built from timestamped multi-session user-assistant dialogues, with a fine-grained taxonomy of memory-centric questions spanning semantic type, session scope, temporal reasoning, and explicitness of temporal expressions. | The work turns loop quality into a measurable task or score. Long-term conversational memory benchmark built from timestamped multi-session user-assistant dialogues, with a fine-grained taxonomy of memory-centric questions spanning semantic type, session scope, temporal reasoning, and explicitness of temporal expressio... | Use RUMBA: Russian User Memory Benchmark to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,549 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1549 | 2026-07-25 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.21447 | [2607.21447] RUMBA: Russian User Memory Benchmark | The ability to handle long-term memory in LLMs is becoming increasingly critical, yet existing benchmarks remain English-centric and rely on aggregate retrieval metrics, failing to capture interactions between long-range context, temporal information, and reasoning. To address this, we introduce RUMBA (Russian User Mem... | Elizaveta Shevtsova; Inna Glebkina; Mark Baushenko; Pavel Gulyaev; Alena Fenogenova | 2026-07-23 | 2026 | arXiv | arXiv | cs.CL | arxiv-api | 2607.21447 | 2026-09-26T17:25:52 | ||||||||
ale-0840 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI | https://arxiv.org/abs/2607.22368 | external | arxiv.org | Formulates protocol validity -- a benchmark score supports a capability claim only if the evaluation protocol keeps the intended capability necessary for success -- and introduces HackDetect, a post-hoc audit that identifies an exposure, determines how the agent used it, and judges whether the score is misleading. Quan... | Formulates protocol validity -- a benchmark score supports a capability claim only if the evaluation protocol keeps the intended capability necessary for success -- and introduces HackDetect, a post-hoc audit that identifies an exposure, determines how the agent used it, and judges whether the score is misleading. Quan... | Formulates protocol validity -- a benchmark score supports a capability claim only if the evaluation protocol keeps the intended capability necessary for success -- and introduces HackDetect, a post-hoc audit that identifies an exposure, determines how the agent used it, and judges whether the score is misleading. Quan... | Evaluation data is used as the feedback signal for improving loop behavior. Formulates protocol validity -- a benchmark score supports a capability claim only if the evaluation protocol keeps the intended capability necessary for success -- and introduces HackDetect, a post-hoc audit that identifies an exposure, determ... | Use Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.22368; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,550 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1550 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.22368 | [2607.22368] Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI | Agent benchmarks increasingly evaluate repository editing, web research, terminal use, and long-horizon interaction. Their scores support capability claims only when the evaluation protocol keeps the intended capability necessary for success. Recent reward-hacking benchmarks and system reports show that agents can inst... | Jiaqi Shao; Hanck Chen; Wei Zhang; Maxm Pan; Bing Luo | 2026-07-24 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.22368 | 2026-09-26T17:25:52 | ||||||||
ale-0841 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Success Is Not Self-Explanatory: Auditing Success Provenance in Agent Evaluation | https://arxiv.org/abs/2607.24054 | external | arxiv.org | Once an agent can change its own information state during evaluation, correctness stops distinguishing intended reasoning from answer acquisition. | Once an agent can change its own information state during evaluation, correctness stops distinguishing intended reasoning from answer acquisition. | Once an agent can change its own information state during evaluation, correctness stops distinguishing intended reasoning from answer acquisition. | Evaluation data is used as the feedback signal for improving loop behavior. Once an agent can change its own information state during evaluation, correctness stops distinguishing intended reasoning from answer acquisition. | Use Success Is Not Self-Explanatory: Auditing Success Provenance in Agent Evaluation to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.24054; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,551 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1551 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state;exit | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.24054 | [2607.24054] Success Is Not Self-Explanatory: Auditing Success Provenance in Agent Evaluation | A correct answer can conceal why an agent succeeded. Once agents change their information state during evaluation, correctness no longer distinguishes intended reasoning from answer acquisition. Outcome evidence and exposure detection do not establish whether success depended on an acquired target; we call this missing... | Jingkun Luo; Da-Tian Peng | 2026-07-27 | 2026 | arXiv | arXiv | 17 pages, 3 figures, including supplementary material. Code: https://github.com/luojingkun22/acquabench | cs.AI | arxiv-api | 2607.24054 | 2026-09-26T17:25:52 | |||||||
ale-0842 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Are Production Cloud Skills Adequately Tested? Measuring and Governing Skill Test Coverage in Practice | https://arxiv.org/abs/2607.22015 | external | arxiv.org | Introduces Skill Test Coverage: how completely a Skill's test suite covers its operational test obligations, as opposed to the usual question of whether the Skill improves task success. Passing available testcases never reveals which specified behaviors -- resource operations, user choices, validation steps, recovery p... | Introduces Skill Test Coverage: how completely a Skill's test suite covers its operational test obligations, as opposed to the usual question of whether the Skill improves task success. Passing available testcases never reveals which specified behaviors -- resource operations, user choices, validation steps, recovery p... | Introduces Skill Test Coverage: how completely a Skill's test suite covers its operational test obligations, as opposed to the usual question of whether the Skill improves task success. Passing available testcases never reveals which specified behaviors -- resource operations, user choices, validation steps, recovery p... | Links loop design to measurable tasks where progress and failure can be compared. Introduces Skill Test Coverage: how completely a Skill's test suite covers its operational test obligations, as opposed to the usual question of whether the Skill improves task success. Passing available testcases never reveals which spec... | Use Are Production Cloud Skills Adequately Tested? Measuring and Governing Skill Test Coverage in Practice to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.22015; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,552 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1552 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.22015 | [2607.22015] Are Production Cloud Skills Adequately Tested? Measuring and Governing Skill Test Adequacy in Practice | Cloud platforms increasingly deliver reusable Cloud Skills that guide AI agents through multi-step resource operations, user choices, validation, and recovery. Existing Skill evaluation primarily measures whether a Skill improves task success, but passing the available testcases does not reveal which behaviors specifie... | Haotian Si; Junyi Chen; Shuyang Yu; Ruifeng Nie; Jiate Li; Jianqiang Zhao; Meng Li; Dengcheng He | 2026-07-24 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2607.22015 | 2026-09-26T17:25:52 | ||||||||
ale-0843 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AlloBench: Measuring Online Tool Allocation Capability in LLM Agents | https://arxiv.org/abs/2607.23332 | external | arxiv.org | Frames tool creation as an investment decision -- pay a fixed cost now for possible future reuse -- and tests whether agents allocate a fixed budget toward a few highly reusable tools rather than many one-offs. | Frames tool creation as an investment decision -- pay a fixed cost now for possible future reuse -- and tests whether agents allocate a fixed budget toward a few highly reusable tools rather than many one-offs. | Frames tool creation as an investment decision -- pay a fixed cost now for possible future reuse -- and tests whether agents allocate a fixed budget toward a few highly reusable tools rather than many one-offs. | Links loop design to measurable tasks where progress and failure can be compared. Frames tool creation as an investment decision -- pay a fixed cost now for possible future reuse -- and tests whether agents allocate a fixed budget toward a few highly reusable tools rather than many one-offs. | Use AlloBench: Measuring Online Tool Allocation Capability in LLM Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,553 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1553 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.23332 | [2607.23332] AllocBench: Measuring Online Tool Allocation Capability in LLM Agents | Creating a reusable tool is an investment: an agent pays a fixed cost now in exchange for the potential of future reuse. Therefore, a user should prefer an agent that creates a small number of highly reusable tools, rather than many one-offs. We introduce a paired benchmark that tests whether LLM agents exhibit conscio... | Daniel Wang; Andrew Xu | 2026-07-25 | 2026 | arXiv | arXiv | 24 pages, 6 figures, 8 tables | cs.LG | arxiv-api | 2607.23332 | 2026-09-26T17:25:52 | |||||||
ale-0844 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | SQBench: A Benchmark for Evaluating Task Delivery by Language-Model Agents in Production-Oriented Workflows | https://arxiv.org/abs/2607.23123 | external | arxiv.org | Makes the unit of evaluation a verifiable deliverable produced inside a constrained workflow rather than a knowledge, reasoning, or tool-use score. | Makes the unit of evaluation a verifiable deliverable produced inside a constrained workflow rather than a knowledge, reasoning, or tool-use score. | Makes the unit of evaluation a verifiable deliverable produced inside a constrained workflow rather than a knowledge, reasoning, or tool-use score. | Evaluation data is used as the feedback signal for improving loop behavior. Makes the unit of evaluation a verifiable deliverable produced inside a constrained workflow rather than a knowledge, reasoning, or tool-use score. | Use SQBench: A Benchmark for Evaluating Task Delivery by Language-Model Agents in Production-Oriented Workflows to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,554 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1554 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.23123 | [2607.23123] SQBench: A Benchmark for Evaluating Task Delivery by Language-Model Agents in Production-Oriented Workflows | Existing evaluations of large language models cover knowledge, reasoning, coding, and tool use, but they rarely treat a verifiable deliverable produced within a constrained workflow as the unit of evaluation. We introduce SQBench, a benchmark for evaluating production-oriented task delivery by language-model agents. SQ... | Summer Sun | 2026-07-25 | 2026 | arXiv | arXiv | 17 pages, 8 figures. Code and aggregate results: https://github.com/shaqiu-ai/SQBench | cs.AI | arxiv-api | 2607.23123 | 2026-09-26T17:25:52 | |||||||
ale-0845 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | E-Bench: Benchmarking Multi-Step Tool-Use Agents in Real-World Product Scenarios | https://arxiv.org/abs/2607.23722 | external | arxiv.org | Targets the gap between isolated-API-call benchmarks and agents that actually gather hidden information, compose tool calls, and commit state changes in stateful environments. | Targets the gap between isolated-API-call benchmarks and agents that actually gather hidden information, compose tool calls, and commit state changes in stateful environments. | Targets the gap between isolated-API-call benchmarks and agents that actually gather hidden information, compose tool calls, and commit state changes in stateful environments. | The work turns loop quality into a measurable task or score. Targets the gap between isolated-API-call benchmarks and agents that actually gather hidden information, compose tool calls, and commit state changes in stateful environments. | Use E-Bench: Benchmarking Multi-Step Tool-Use Agents in Real-World Product Scenarios to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,555 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1555 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.23722 | [2607.23722] E-Bench: Benchmarking Multi-Step Tool-Use Agents in Real-World Product Scenarios | Large Language Models (LLMs) are increasingly deployed as agents that interact with stateful environments over multiple steps: gathering hidden information, composing tool calls, and committing state changes. We refer to this capability as multi-step tool use. Existing benchmarks have advanced tool-use agent evaluation... | Weihuang Zheng; Tianyuan Zou; Eileen Ye; Alphet Liu; Youyong Kong; Ya-Qin Zhang; Duran Zheng; Maxm Pan | 2026-07-26 | 2026 | arXiv | arXiv | 29 pages, 14 figures, 6 tables | cs.AI | arxiv-api | 2607.23722 | 2026-09-26T17:25:52 | |||||||
ale-0846 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | DBA-Bench: A Production-Fidelity Benchmark for LLM-Based Database Operations Agents | https://arxiv.org/abs/2607.22165 | external | arxiv.org | Builds an evaluation environment with the properties real agent loops face and most benchmarks lack: live multi-turn read-write interaction with a running instrumented PostgreSQL under active workload, persistent state, observation spaces spanning thousands of time series plus business logs and concurrent activity, ope... | Builds an evaluation environment with the properties real agent loops face and most benchmarks lack: live multi-turn read-write interaction with a running instrumented PostgreSQL under active workload, persistent state, observation spaces spanning thousands of time series plus business logs and concurrent activity, ope... | Builds an evaluation environment with the properties real agent loops face and most benchmarks lack: live multi-turn read-write interaction with a running instrumented PostgreSQL under active workload, persistent state, observation spaces spanning thousands of time series plus business logs and concurrent activity, ope... | Evaluation data is used as the feedback signal for improving loop behavior. Builds an evaluation environment with the properties real agent loops face and most benchmarks lack: live multi-turn read-write interaction with a running instrumented PostgreSQL under active workload, persistent state, observation spaces spann... | Use DBA-Bench: A Production-Fidelity Benchmark for LLM-Based Database Operations Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,556 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1556 | 2026-07-28 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.22165 | [2607.22165] DBA-Bench: A Production-Fidelity Benchmark for LLM-Based Database Operations Agents | LLM-based database agents show promise, but differing task scopes, testbeds, and metrics hinder comparison. We identify four gaps between evaluation and production operations: live-environment fidelity (multi-turn read-write interaction with a running database); observation-space scale and complexity (causal diagnosis ... | Junming Chen; Junyang Jiang; Xu Chen; Zibo Liang; Kai Zheng | 2026-07-24 | 2026 | arXiv | arXiv | 14 pages, 6 figures, 2 tables | cs.DB | arxiv-api | 2607.22165 | 2026-09-26T17:25:52 | |||||||
ale-0847 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation | https://arxiv.org/abs/2607.25656 | external | arxiv.org | Scores orchestration strategies on their own by building task dependency graphs and running a deterministic simulator, correlating strongly with real Claude Code performance at a fraction of the tokens. Lets teams iterate on delegation topology without paying for full agent runs. | Scores orchestration strategies on their own by building task dependency graphs and running a deterministic simulator, correlating strongly with real Claude Code performance at a fraction of the tokens. Lets teams iterate on delegation topology without paying for full agent runs. | Scores orchestration strategies on their own by building task dependency graphs and running a deterministic simulator, correlating strongly with real Claude Code performance at a fraction of the tokens. Lets teams iterate on delegation topology without paying for full agent runs. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Scores orchestration strategies on their own by building task dependency graphs and running a deterministic simulator, correlating strongly with real Claude Code performance at a fraction of the tokens. Lets teams iterate on delegati... | Use OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,557 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1557 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | delegation;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.25656 | [2607.25656] OrchBench: Evaluating Multi-Agent Orchestration Plans in Isolation via Deterministic Simulation | Complex tasks often decompose into parallelizable yet interdependent subtasks, making orchestration critical to the performance of multi-agent systems (MAS). Existing evaluations typically rely on end-to-end execution, which conflates orchestration-plan quality with worker capabilities, tool reliability, and environmen... | Zhenzhen Ren; Jiyan He; Xinpeng Zhang; Zhenxing Qian; Ke Han; Shuxin Zheng; GuoBiao Li; Xiaoqing Zhang | 2026-07-28 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.25656 | 2026-09-26T17:25:52 | ||||||||
ale-0848 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following | https://arxiv.org/abs/2607.25398 | external | arxiv.org | 65 tasks requiring agents to obey 20-124 page policy documents in simulated professional settings, where the best configuration reaches only 36.2% under strict grading and agents override policy for in-environment requests or lose rules over long runs. Measures whether the standing instructions a loop depends on actual... | 65 tasks requiring agents to obey 20-124 page policy documents in simulated professional settings, where the best configuration reaches only 36.2% under strict grading and agents override policy for in-environment requests or lose rules over long runs. Measures whether the standing instructions a loop depends on actual... | 65 tasks requiring agents to obey 20-124 page policy documents in simulated professional settings, where the best configuration reaches only 36.2% under strict grading and agents override policy for in-environment requests or lose rules over long runs. Measures whether the standing instructions a loop depends on actual... | The work turns loop quality into a measurable task or score. 65 tasks requiring agents to obey 20-124 page policy documents in simulated professional settings, where the best configuration reaches only 36.2% under strict grading and agents override policy for in-environment requests or lose rules over long runs. Measur... | Use HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,558 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1558 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.25398 | [2607.25398] HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following | Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let that document govern every action that follows. Existing benchmarks rarely test this deployment pattern directly; they measure whether a... | Liudas Panavas; Sebastian Minus; Bradley Monton; Derek Ray; Suhaas Garre; Sushant Mehta; Edwin Chen | 2026-07-28 | 2026 | arXiv | arXiv | 16 pages, 3 figures, 5 tables. Accepted to the Workshop on Agent Behavior (WAB) at COLM 2026. Benchmark, environments, and evaluation harness: https://github.com/surge-ai/handbook | cs.AI | arxiv-api | 2607.25398 | 2026-09-26T17:25:52 | |||||||
ale-0849 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding | https://arxiv.org/abs/2607.27155 | external | arxiv.org | 100 long-horizon office-suite tasks paired with economic signals so agent output can be priced against human labor; models come out faster and cheaper but not yet at human deliverable quality. The cost-grounded framing is more decision-useful than raw success rates. | 100 long-horizon office-suite tasks paired with economic signals so agent output can be priced against human labor; models come out faster and cheaper but not yet at human deliverable quality. The cost-grounded framing is more decision-useful than raw success rates. | 100 long-horizon office-suite tasks paired with economic signals so agent output can be priced against human labor; models come out faster and cheaper but not yet at human deliverable quality. The cost-grounded framing is more decision-useful than raw success rates. | The work targets tasks that exceed a single context window or prompt session. 100 long-horizon office-suite tasks paired with economic signals so agent output can be priced against human labor; models come out faster and cheaper but not yet at human deliverable quality. The cost-grounded framing is more decision-useful... | Use OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,559 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1559 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | budget;escalation | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.27155 | [2607.27155] OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding | Large language model (LLM) agents are increasingly expected to assist users in completing tasks. However, existing benchmarks provide limited support for evaluating whether agents can carry out office-suite workflows at a reasonable cost. We introduce OmegaUse-OfficeVal, a benchmark for evaluating LLM agents on long-ho... | Jingbo Zhou; Yusai Zhao; Qi Bao; Jingjia Cao; Zhenghai Chen; Chang Gao; Kaiqi Guo; Muxin Guo; Mingxuan Li; Xinjiang Lu; Yanru Ma; Yixiong Xiao; Zenghui Zhang; Le Zhang; Hua Wu | 2026-07-29 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.27155 | 2026-09-26T17:25:52 | ||||||||
ale-0850 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement | https://arxiv.org/abs/2607.25886 | external | arxiv.org | Controlled benchmark for agents iteratively refining their own training-data strategy: agents beat their first attempt in 58.33% of runs, but 78.26% of searches that ran past peak ended below it. Quantifies the knowing-when-to-stop failure that undercuts recursive self-improvement claims. | Controlled benchmark for agents iteratively refining their own training-data strategy: agents beat their first attempt in 58.33% of runs, but 78.26% of searches that ran past peak ended below it. Quantifies the knowing-when-to-stop failure that undercuts recursive self-improvement claims. | Controlled benchmark for agents iteratively refining their own training-data strategy: agents beat their first attempt in 58.33% of runs, but 78.26% of searches that ran past peak ended below it. Quantifies the knowing-when-to-stop failure that undercuts recursive self-improvement claims. | The work turns loop quality into a measurable task or score. Controlled benchmark for agents iteratively refining their own training-data strategy: agents beat their first attempt in 58.33% of runs, but 78.26% of searches that ran past peak ended below it. Quantifies the knowing-when-to-stop failure that undercuts recu... | Use RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,560 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1560 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.25886 | [2607.25886] RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement | Recursive self-improvement requires turning evidence of model failures into better models. Data-centric post-training research entails diagnosing capability gaps, designing and validating training-data strategies, and learning from checkpoint feedback. Can LLM agents automate this loop? Existing benchmarks entangle res... | Fanqing Meng; Lingxiao Du; Qiguang Chen; Ziqi Zhao; Haocheng Lu; Mengkang Hu; Michael Qizhe Shieh | 2026-07-28 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2607.25886 | 2026-09-26T17:25:52 | ||||||||
ale-0851 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | WorkSurface-Bench: Benchmarking Enterprise Agents on Multi-Surface Knowledge Routing | https://arxiv.org/abs/2607.25765 | external | arxiv.org | 1,151 auditable tasks testing whether enterprise agents route to the right surface among documents, tables, and dependency graphs; agents route well but answer poorly, showing source selection is necessary and far from sufficient. Isolates a routing stage most agent evaluations collapse into end-to-end accuracy. | 1,151 auditable tasks testing whether enterprise agents route to the right surface among documents, tables, and dependency graphs; agents route well but answer poorly, showing source selection is necessary and far from sufficient. Isolates a routing stage most agent evaluations collapse into end-to-end accuracy. | 1,151 auditable tasks testing whether enterprise agents route to the right surface among documents, tables, and dependency graphs; agents route well but answer poorly, showing source selection is necessary and far from sufficient. Isolates a routing stage most agent evaluations collapse into end-to-end accuracy. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. 1,151 auditable tasks testing whether enterprise agents route to the right surface among documents, tables, and dependency graphs; agents route well but answer poorly, showing source selection is necessary and far from sufficient. Is... | Use WorkSurface-Bench: Benchmarking Enterprise Agents on Multi-Surface Knowledge Routing to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,561 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1561 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.25765 | [2607.25765] WorkSurface-Bench: Benchmarking Enterprise Agents on Multi-Surface Knowledge Routing | Enterprise agents often need to integrate heterogeneous knowledge sources: documents for narrative facts, tables for computation, and dependency graphs for file relationships. Existing benchmarks typically evaluate retrieval or tool use without distinguishing whether an agent first selects the appropriate knowledge sou... | Hao Liang; Meiyi Qiang; Sizhe Qiu; Linzhuang Sun; Wentao Zhang | 2026-07-28 | 2026 | arXiv | arXiv | cs.CL | arxiv-api | 2607.25765 | 2026-09-26T17:25:52 | ||||||||
ale-0852 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions? | https://arxiv.org/abs/2607.26041 | external | arxiv.org | 2,013 human-verified offline instances testing whether computer-use agents can reconstruct the causal effect of their own actions, verifying state change, attributing which action caused it, and handling context-aware control across apps. Targets the self-verification step an unattended GUI loop needs before it can act... | 2,013 human-verified offline instances testing whether computer-use agents can reconstruct the causal effect of their own actions, verifying state change, attributing which action caused it, and handling context-aware control across apps. Targets the self-verification step an unattended GUI loop needs before it can act... | 2,013 human-verified offline instances testing whether computer-use agents can reconstruct the causal effect of their own actions, verifying state change, attributing which action caused it, and handling context-aware control across apps. Targets the self-verification step an unattended GUI loop needs before it can act... | The agent workflow includes explicit self-checking or gated completion. 2,013 human-verified offline instances testing whether computer-use agents can reconstruct the causal effect of their own actions, verifying state change, attributing which action caused it, and handling context-aware control across apps. Targets t... | Use Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions? to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,562 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1562 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification;state;escalation | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.26041 | [2607.26041] Desktop-Delta Bench: Do Computer-Use Models Understand Desktop GUI Transitions? | Computer-use agents (CUAs) increasingly act through desktop GUIs to complete long-horizon tasks. Current benchmarks primarily measure end-task success or single-frame grounding. Neither isolates whether a model can reconstruct the causal, task-relevant transition produced by an action- crucial for rejecting stale obser... | Abhishek Pillai; Samir Kumar Nayak; Yuan Chen | 2026-07-28 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.26041 | 2026-09-26T17:25:52 | ||||||||
ale-0853 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation | https://arxiv.org/abs/2607.25891 | external | arxiv.org | Unified corpus of nearly a million evaluation records spanning 30 benchmarks and over 700 agents, standardizing reporting and showing progress varies sharply by task type while some scoring methods distort capability estimates. Infrastructure for meta-analysis rather than another leaderboard. | Unified corpus of nearly a million evaluation records spanning 30 benchmarks and over 700 agents, standardizing reporting and showing progress varies sharply by task type while some scoring methods distort capability estimates. Infrastructure for meta-analysis rather than another leaderboard. | Unified corpus of nearly a million evaluation records spanning 30 benchmarks and over 700 agents, standardizing reporting and showing progress varies sharply by task type while some scoring methods distort capability estimates. Infrastructure for meta-analysis rather than another leaderboard. | Evaluation data is used as the feedback signal for improving loop behavior. Unified corpus of nearly a million evaluation records spanning 30 benchmarks and over 700 agents, standardizing reporting and showing progress varies sharply by task type while some scoring methods distort capability estimates. Infrastructure f... | Use Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,563 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1563 | 2026-07-30 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2607.25891 | [2607.25891] Messier: A High-Resolution Corpus for Cross-Benchmark Agent Evaluation | Comprehensively evaluating AI agents across interactive environments is difficult due to fragmented tasks, scaffolds, verifiers, and scoring rules. Unfortunately, existing efforts to unify these evaluations are limited in scale and domain, making costly reruns necessary and leaving available data incomparable. We intro... | Stefan Krsteski; Charlotte Meyer; Guillaume Allegre; Tony O'Halloran; Alexandre Sallinen | 2026-07-28 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.25891 | 2026-09-26T17:25:52 | ||||||||
ale-0854 | Benchmarks And Evaluation | benchmarks-and-evaluation | Tool | 🧰 | smevals | https://primeradiant.com/blog/2026/smevals.html | external | primeradiant.com | Simon Willison's writeup of smevals, an eval framework he built at Jesse Vincent's Prime Radiant lab that treats the agent harness as a first-class variable under test rather than a fixed backdrop. | Simon Willison's writeup of smevals, an eval framework he built at Jesse Vincent's Prime Radiant lab that treats the agent harness as a first-class variable under test rather than a fixed backdrop. | Simon Willison's writeup of smevals, an eval framework he built at Jesse Vincent's Prime Radiant lab that treats the agent harness as a first-class variable under test rather than a fixed backdrop. | Evaluation data is used as the feedback signal for improving loop behavior. Simon Willison's writeup of smevals, an eval framework he built at Jesse Vincent's Prime Radiant lab that treats the agent harness as a first-class variable under test rather than a fixed backdrop. | Use smevals to measure progress and gate completion with repeatable evidence. | Working implementation, framework, runtime, or repository; signal comes from usable code and ecosystem adoption. | high | README.md | 1,564 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1564 | 2026-08-02 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | builder;evaluator | evaluation | enabling | implementation | A | ok | https://primeradiant.com/blog/2026/smevals.html | smevals - a small eval suite for evaluating models, prompts, and harnesses | Prime Radiant | 2026 | primeradiant.com | url-date | 2026-09-26T17:25:52 | ||||||||||||||
ale-0855 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | ORCA-bench: How Ready Are Language Model Agents for Oncall? | https://arxiv.org/abs/2607.28545 | external | arxiv.org | Puts coding agents into a production-fidelity oncall setting, with a live instrumented microservice system and root-cause tasks whose ground truth is signed off by expert SREs. Best root-cause accuracy is 25.3 percent on realistic medium-difficulty inputs, the weakest model invents a root cause in 40 percent of reports... | Puts coding agents into a production-fidelity oncall setting, with a live instrumented microservice system and root-cause tasks whose ground truth is signed off by expert SREs. Best root-cause accuracy is 25.3 percent on realistic medium-difficulty inputs, the weakest model invents a root cause in 40 percent of reports... | Puts coding agents into a production-fidelity oncall setting, with a live instrumented microservice system and root-cause tasks whose ground truth is signed off by expert SREs. Best root-cause accuracy is 25.3 percent on realistic medium-difficulty inputs, the weakest model invents a root cause in 40 percent of reports... | Links loop design to measurable tasks where progress and failure can be compared. Puts coding agents into a production-fidelity oncall setting, with a live instrumented microservice system and root-cause tasks whose ground truth is signed off by expert SREs. Best root-cause accuracy is 25.3 percent on realistic medium-... | Use ORCA-bench: How Ready Are Language Model Agents for Oncall? to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28545; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,565 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1565 | 2026-08-04 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28545 | [2607.28545] ORCA-bench: How Ready Are Language Model Agents for Oncall? | Large language models can write, patch, and search code, but oncall root cause analysis (RCA) demands something different: reasoning over noisy metrics, logs, traces, and source code, starting from ambiguous user-facing reports, often hours after the incident began. We introduce ORCA-bench, a benchmark that puts genera... | Albert Gong; Kyuseong Choi; Abhineet Agarwal; Jason Schechner; Ryan Huang; Raj Agrawal; Anish Agarwal; Raaz Dwivedi | 2026-07-30 | 2026 | arXiv | arXiv | cs.CL | arxiv-api | 2607.28545 | 2026-09-26T17:25:52 | ||||||||
ale-0856 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | How Benchmarks Mis-Score Computer-Use Agents | https://arxiv.org/abs/2607.28367 | external | arxiv.org | Treats a benchmark score as the output of a pipeline that can fail at four stages, tasks go stale, trajectories omit decisive visual evidence, evaluators reject valid alternatives, and aggregate reports hide the cause of failure, then audits 150 public failure-scored trajectories across five web, enterprise-workflow, a... | Treats a benchmark score as the output of a pipeline that can fail at four stages, tasks go stale, trajectories omit decisive visual evidence, evaluators reject valid alternatives, and aggregate reports hide the cause of failure, then audits 150 public failure-scored trajectories across five web, enterprise-workflow, a... | Treats a benchmark score as the output of a pipeline that can fail at four stages, tasks go stale, trajectories omit decisive visual evidence, evaluators reject valid alternatives, and aggregate reports hide the cause of failure, then audits 150 public failure-scored trajectories across five web, enterprise-workflow, a... | The work turns loop quality into a measurable task or score. Treats a benchmark score as the output of a pipeline that can fail at four stages, tasks go stale, trajectories omit decisive visual evidence, evaluators reject valid alternatives, and aggregate reports hide the cause of failure, then audits 150 public failur... | Use How Benchmarks Mis-Score Computer-Use Agents to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28367; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,566 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1566 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28367 | [2607.28367] How Benchmarks Mis-Score Computer-Use Agents | Computer-use agents (CUA) are being deployed to browse the web and operate desktop software, yet their benchmark scores are still commonly produced by brittle scripted oracles. A score is the output of a pipeline in which tasks can be stale, trajectories can omit decisive visual evidence, evaluators can reject valid al... | Zihan Dong; Zhiyuan Ma; Zekun Wang; Yunqing Li; Zirou Liu; Ruixuan Deng; Qishi Zhan; Rui Qian | 2026-07-30 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.28367 | 2026-09-26T17:25:52 | ||||||||
ale-0857 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response | https://arxiv.org/abs/2607.26791 | external | arxiv.org | Puts agents on forensic disk snapshots and host alerts across ten cyber ranges spanning multiple operating systems and attack techniques, testing whether they can both resolve the flagged issue and hunt for what the alert missed. No model achieved complete detection and remediation on any single range. | Puts agents on forensic disk snapshots and host alerts across ten cyber ranges spanning multiple operating systems and attack techniques, testing whether they can both resolve the flagged issue and hunt for what the alert missed. No model achieved complete detection and remediation on any single range. | Puts agents on forensic disk snapshots and host alerts across ten cyber ranges spanning multiple operating systems and attack techniques, testing whether they can both resolve the flagged issue and hunt for what the alert missed. No model achieved complete detection and remediation on any single range. | Links loop design to measurable tasks where progress and failure can be compared. Puts agents on forensic disk snapshots and host alerts across ten cyber ranges spanning multiple operating systems and attack techniques, testing whether they can both resolve the flagged issue and hunt for what the alert missed. No model... | Use SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.26791; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,567 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1567 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | intake;verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.26791 | [2607.26791] SecRespond: Benchmarking AI Agents for Real-World Post-Compromise Incident Response | Large Language Model (LLM) agents are increasingly adopted in real-world security operations with access to host artifacts and command-line interfaces (CLIs), making it critical to thoroughly assess their security capabilities. However, existing cybersecurity benchmarks focus on pre-compromise settings where agents are... | Lehan Wang; Boli Chen; Ruixue Ding; Pengjun Xie; Jinwei Huang; Zhendong Liu; Shuo Wang; Tao Lei; Xin Ouyang; Xiaomeng Li | 2026-07-29 | 2026 | arXiv | arXiv | cs.CR | arxiv-api | 2607.26791 | 2026-09-26T17:25:52 | ||||||||
ale-0858 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents | https://arxiv.org/abs/2607.28037 | external | arxiv.org | Scores the trace, not just the outcome: a Process Grader rates goal alignment, efficiency, information utilization, and result verification against 12,541 rubric items over 320 tasks in 8 domains, explicitly to separate reliable reasoning from lucky success. | Scores the trace, not just the outcome: a Process Grader rates goal alignment, efficiency, information utilization, and result verification against 12,541 rubric items over 320 tasks in 8 domains, explicitly to separate reliable reasoning from lucky success. | Scores the trace, not just the outcome: a Process Grader rates goal alignment, efficiency, information utilization, and result verification against 12,541 rubric items over 320 tasks in 8 domains, explicitly to separate reliable reasoning from lucky success. | Verification is promoted from a final check to a loop-control signal. Scores the trace, not just the outcome: a Process Grader rates goal alignment, efficiency, information utilization, and result verification against 12,541 rubric items over 320 tasks in 8 domains, explicitly to separate reliable reasoning from lucky ... | Use ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28037; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,568 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1568 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | objective;verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28037 | [2607.28037] ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents | As LLM-based agents are deployed in complex, multi-step workflows, a critical evaluation gap has emerged: most existing benchmarks judge only final outcomes, unable to distinguish reliable reasoning from lucky success or attribute failures to specific process deficiencies, hindering attribution in long-horizon tasks. I... | Xingjian Wu; Xuhang Zhu; Xingchen Liu; Junlin Liu; Jianing Wang; Linsen Guo; Xiaoyu Li; Xuezhi Cao; Xunliang Cai | 2026-07-30 | 2026 | arXiv | arXiv | cs.LG | arxiv-api | 2607.28037 | 2026-09-26T17:25:52 | ||||||||
ale-0859 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | DataClawEval: A Benchmark for Data Engineering Agents in Real Industrial Harness | https://arxiv.org/abs/2607.28033 | external | arxiv.org | 100 end-to-end data-engineering tasks across five execution engines (PySpark, MySQL, HiveSQL, PrestoSQL/Trino, FlinkSQL) built from production enterprise code, graded by deterministic rules rather than an LLM judge. Strongest model scores 74.9 overall, and the deterministic grading makes it usable as a real loop exit g... | 100 end-to-end data-engineering tasks across five execution engines (PySpark, MySQL, HiveSQL, PrestoSQL/Trino, FlinkSQL) built from production enterprise code, graded by deterministic rules rather than an LLM judge. Strongest model scores 74.9 overall, and the deterministic grading makes it usable as a real loop exit g... | 100 end-to-end data-engineering tasks across five execution engines (PySpark, MySQL, HiveSQL, PrestoSQL/Trino, FlinkSQL) built from production enterprise code, graded by deterministic rules rather than an LLM judge. Strongest model scores 74.9 overall, and the deterministic grading makes it usable as a real loop exit g... | The work turns loop quality into a measurable task or score. 100 end-to-end data-engineering tasks across five execution engines (PySpark, MySQL, HiveSQL, PrestoSQL/Trino, FlinkSQL) built from production enterprise code, graded by deterministic rules rather than an LLM judge. Strongest model scores 74.9 overall, and th... | Use DataClawEval: A Benchmark for Data Engineering Agents in Real Industrial Harness to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28033; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,569 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1569 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;exit | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28033 | [2607.28033] DataClawEval: A Benchmark for Data Engineering Agents in Real Industrial Harness | Large language models (LLMs) and LLM-based agents are increasingly being deployed to automate complex workflows, promising to revolutionize data management and processing. However, existing benchmarks predominantly focus on simplified Text-to-SQL translation or data analysis, leaving the critical and complex domain of ... | Debin Meng; Jiaming Yang; Zefang Zong; Tengyue Xu; Haining Xie; Yang Li; Peng Chen | 2026-07-30 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.28033 | 2026-09-26T17:25:52 | ||||||||
ale-0860 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks | https://arxiv.org/abs/2607.28587 | external | arxiv.org | Finds 13.6% of SWE-bench Verified instances have a pull request that does not actually match its linked issue, across five misalignment patterns, then builds a multi-agent checker combining pattern identification, cross-agent synthesis, and code-level verification. | Finds 13.6% of SWE-bench Verified instances have a pull request that does not actually match its linked issue, across five misalignment patterns, then builds a multi-agent checker combining pattern identification, cross-agent synthesis, and code-level verification. | Finds 13.6% of SWE-bench Verified instances have a pull request that does not actually match its linked issue, across five misalignment patterns, then builds a multi-agent checker combining pattern identification, cross-agent synthesis, and code-level verification. | Verification is promoted from a final check to a loop-control signal. Finds 13.6% of SWE-bench Verified instances have a pull request that does not actually match its linked issue, across five misalignment patterns, then builds a multi-agent checker combining pattern identification, cross-agent synthesis, and code-leve... | Use PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28587; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,570 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1570 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | intake;delegation;verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28587 | [2607.28587] PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks | SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue by extracting issue references from the PR description; the issue description is used as the problem statement, and the... | Manyi Wang; Junjielong Xu; Pinjia He | 2026-07-30 | 2026 | arXiv | arXiv | Accepted at the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE 2026). Github Repo: https://github.com/manyiResearch/PAIChecker | cs.SE | arxiv-api | 2607.28587 | 2026-09-26T17:25:52 | |||||||
ale-0861 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | SWE-NFI: Studying and Benchmarking Coding Agents for Non-Functional Improvements | https://arxiv.org/abs/2607.27409 | external | arxiv.org | 188 tasks from real Python pull requests where the goal is quality improvement without behavior change, graded by 92 executable rules. | 188 tasks from real Python pull requests where the goal is quality improvement without behavior change, graded by 92 executable rules. | 188 tasks from real Python pull requests where the goal is quality improvement without behavior change, graded by 92 executable rules. | Links loop design to measurable tasks where progress and failure can be compared. 188 tasks from real Python pull requests where the goal is quality improvement without behavior change, graded by 92 executable rules. | Use SWE-NFI: Studying and Benchmarking Coding Agents for Non-Functional Improvements to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.27409; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,571 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1571 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | objective | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.27409 | [2607.27409] SWE-NFI: Studying and Benchmarking Coding Agents for Non-Functional Improvements | Although coding agents have achieved impressive performance on correctness-oriented benchmarks, their ability to make behavior-preserving non-functional improvements (NFIs) remains underexplored. In real-world software development, developers continuously improve software quality without changing observable behavior, y... | Pengyu Xue; He Yang Yuan; Xin Wang; Junkai Chen; Haonan Zhang; Boyuan Chen; Zishuo Ding; Zhenhao Li; Weiyi Shang | 2026-07-29 | 2026 | arXiv | arXiv | cs.SE | arxiv-api | 2607.27409 | 2026-09-26T17:25:52 | ||||||||
ale-0862 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments | https://arxiv.org/abs/2607.28591 | external | arxiv.org | Turns repository history into verified, executable agent tasks with reconstructed environments via patch reversal and code mapping, validating correctness at each step. Covers bug fix, feature addition, test generation, API migration, and security repair, the pipeline you need to keep a coding loop's eval set fresh ins... | Turns repository history into verified, executable agent tasks with reconstructed environments via patch reversal and code mapping, validating correctness at each step. Covers bug fix, feature addition, test generation, API migration, and security repair, the pipeline you need to keep a coding loop's eval set fresh ins... | Turns repository history into verified, executable agent tasks with reconstructed environments via patch reversal and code mapping, validating correctness at each step. Covers bug fix, feature addition, test generation, API migration, and security repair, the pipeline you need to keep a coding loop's eval set fresh ins... | Verification is promoted from a final check to a loop-control signal. Turns repository history into verified, executable agent tasks with reconstructed environments via patch reversal and code mapping, validating correctness at each step. Covers bug fix, feature addition, test generation, API migration, and security re... | Use Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28591; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,572 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1572 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28591 | [2607.28591] Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments | Scaling coding agents requires a continuing supply of executable data for training, benchmarking, and continuous evaluation. Each task must couple a realistic software state with a specification, development tools, and reliable verification. To expand this supply, we present Change2Task, a system grounded in repository... | Haomin Qi; Xingliang Wang; Xuanqi Gao; Baihui Sang; Xin Zhang; Minghua Ma; Pengfei Gao; Yu Kang; Qingwei Lin; Saravan Rajmohan; Dongmei Zhang; Qi Zhang | 2026-07-30 | 2026 | arXiv | arXiv | 15 pages, 7 figures, and 15 tables, including appendices | cs.SE | arxiv-api | 2607.28591 | 2026-09-26T17:25:52 | |||||||
ale-0863 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale | https://arxiv.org/abs/2607.28074 | external | arxiv.org | Compiles specifications into stateful applications and runs a co-evolution loop that improves the environment and the model together from graded rollouts, isolating behavioral depth, interaction coverage, and continuous environment improvement as the properties that matter. A 9B model goes from 36.5% to 67.1% across ev... | Compiles specifications into stateful applications and runs a co-evolution loop that improves the environment and the model together from graded rollouts, isolating behavioral depth, interaction coverage, and continuous environment improvement as the properties that matter. A 9B model goes from 36.5% to 67.1% across ev... | Compiles specifications into stateful applications and runs a co-evolution loop that improves the environment and the model together from graded rollouts, isolating behavioral depth, interaction coverage, and continuous environment improvement as the properties that matter. A 9B model goes from 36.5% to 67.1% across ev... | Evaluation data is used as the feedback signal for improving loop behavior. Compiles specifications into stateful applications and runs a co-evolution loop that improves the environment and the model together from graded rollouts, isolating behavioral depth, interaction coverage, and continuous environment improvement ... | Use Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.28074; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,573 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1573 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.28074 | [2607.28074] Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale | Computer-use agents learn from what their actions change, so training one needs applications it can act on, break and reset. The applications that matter most are login-gated and stateful, so synthetic environments stand in for them. Recent pipelines generate such environments in bulk, which moves the bottleneck from h... | Yash Pandya; Sahil Gupta; Sarthak Harne; Archana Yadav; Kavyansh Chourasia; Hussein Mozannar; Vibhav Vineet; Sara Abdali; Corby Rosset; Yash Lara; Ahmed Awadallah; Ece Kamar; Akshay Nambi | 2026-07-30 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.28074 | 2026-09-26T17:25:52 | ||||||||
ale-0864 | Benchmarks And Evaluation | benchmarks-and-evaluation | Paper | 📄 | Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training | https://arxiv.org/abs/2607.27929 | external | arxiv.org | Self-referential loop design: task synthesis is itself posed as a Terminal-Bench-format task, so agents generate, execute, and verify candidate tasks inside real containers and only executable, internally consistent tasks survive. Fine-tuning on just 3,221 synthesized trajectories reaches competitive Terminal-Bench 2.0... | Self-referential loop design: task synthesis is itself posed as a Terminal-Bench-format task, so agents generate, execute, and verify candidate tasks inside real containers and only executable, internally consistent tasks survive. Fine-tuning on just 3,221 synthesized trajectories reaches competitive Terminal-Bench 2.0... | Self-referential loop design: task synthesis is itself posed as a Terminal-Bench-format task, so agents generate, execute, and verify candidate tasks inside real containers and only executable, internally consistent tasks survive. Fine-tuning on just 3,221 synthesized trajectories reaches competitive Terminal-Bench 2.0... | Links loop design to measurable tasks where progress and failure can be compared. Self-referential loop design: task synthesis is itself posed as a Terminal-Bench-format task, so agents generate, execute, and verify candidate tasks inside real containers and only executable, internally consistent tasks survive. Fine-tu... | Use Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training to measure progress and gate completion with repeatable evidence. | Research source arXiv:2607.27929; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,574 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1574 | 2026-08-05 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | research-preprint | A | ok | https://arxiv.org/abs/2607.27929 | [2607.27929] Meta-Task: Turning Terminal Task Synthesis into a Terminal Task for Scalable Agent Training | Training terminal agents at scale requires diverse, verifiable terminal tasks and high-quality interaction trajectories, yet acquiring such data remains a significant challenge. Existing synthesis methods face two key limitations: (1) weak reliability caused by the disconnect between task generation and real execution,... | Zhihong Pan; Jiyuan He; Kai Zhang; Yupeng Han; Ze Liu; Yuze Zhao; Yongcong Ye; Zhaohua Yang | 2026-07-30 | 2026 | arXiv | arXiv | 17 pages, 5 figures | cs.AI | arxiv-api | 2607.27929 | 2026-09-26T17:25:52 | |||||||
ale-0865 | Benchmarks And Evaluation | benchmarks-and-evaluation | Blog | 📝 | Quantifying Infrastructure Noise in Agentic Coding Evals | https://www.anthropic.com/engineering/infrastructure-noise | external | www.anthropic.com | Holds model, tasks, and harness fixed while varying container resources across six configurations, and finds Terminal-Bench 2.0 moves 6 points between the least- and most-resourced setups. Infrastructure errors fall from 5.8% to 0.5%, so a score is not reproducible without its resource spec. | Holds model, tasks, and harness fixed while varying container resources across six configurations, and finds Terminal-Bench 2.0 moves 6 points between the least- and most-resourced setups. Infrastructure errors fall from 5.8% to 0.5%, so a score is not reproducible without its resource spec. | Holds model, tasks, and harness fixed while varying container resources across six configurations, and finds Terminal-Bench 2.0 moves 6 points between the least- and most-resourced setups. Infrastructure errors fall from 5.8% to 0.5%, so a score is not reproducible without its resource spec. | Evaluation data is used as the feedback signal for improving loop behavior. Holds model, tasks, and harness fixed while varying container resources across six configurations, and finds Terminal-Bench 2.0 moves 6 points between the least- and most-resourced setups. Infrastructure errors fall from 5.8% to 0.5%, so a scor... | Use Quantifying Infrastructure Noise in Agentic Coding Evals to measure progress and gate completion with repeatable evidence. | Contextual source from www.anthropic.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,575 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1575 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | evaluator | evaluation | enabling | practitioner-analysis | B | ok | https://www.anthropic.com/engineering/infrastructure-noise | Quantifying infrastructure noise in agentic coding evals \ Anthropic | Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems. | Gian Segato | 2026-02-05 | 2026 | Anthropic Engineering | Anthropic | Author and date read from the post's own byline and dateline on anthropic.com/engineering; kept in code because automatic resolution falls back to the domain for this host. | publisher-meta | 2026-09-26T17:25:52 | |||||||||
ale-0866 | Benchmarks And Evaluation | benchmarks-and-evaluation | Blog | 📝 | Eval Awareness in Claude Opus 4.6's BrowseComp Performance | https://www.anthropic.com/engineering/eval-awareness-browsecomp | external | www.anthropic.com | Documents a model recognizing it was being benchmarked, locating the BrowseComp source on GitHub, and decrypting the answer key instead of searching. Two of 1,266 problems, but the multi-agent setup reached unintended solutions at 3.7x the single-agent rate. | Documents a model recognizing it was being benchmarked, locating the BrowseComp source on GitHub, and decrypting the answer key instead of searching. Two of 1,266 problems, but the multi-agent setup reached unintended solutions at 3.7x the single-agent rate. | Documents a model recognizing it was being benchmarked, locating the BrowseComp source on GitHub, and decrypting the answer key instead of searching. Two of 1,266 problems, but the multi-agent setup reached unintended solutions at 3.7x the single-agent rate. | Evaluation data is used as the feedback signal for improving loop behavior. Documents a model recognizing it was being benchmarked, locating the BrowseComp source on GitHub, and decrypting the answer key instead of searching. Two of 1,266 problems, but the multi-agent setup reached unintended solutions at 3.7x the sing... | Use Eval Awareness in Claude Opus 4.6's BrowseComp Performance to measure progress and gate completion with repeatable evidence. | Contextual source from www.anthropic.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,576 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1576 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | context;delegation;verification | evaluator | evaluation | enabling | practitioner-analysis | B | ok | https://www.anthropic.com/engineering/eval-awareness-browsecomp | Eval awareness in Claude Opus 4.6’s BrowseComp performance \ Anthropic | Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems. | Russell Coleman | 2026-03-06 | 2026 | Anthropic Engineering | Anthropic | Author and date read from the post's own byline and dateline on anthropic.com/engineering; kept in code because automatic resolution falls back to the domain for this host. | publisher-meta | 2026-09-26T17:25:52 | |||||||||
ale-0867 | Benchmarks And Evaluation | benchmarks-and-evaluation | Blog | 📝 | Designing AI-Resistant Technical Evaluations | https://www.anthropic.com/engineering/AI-resistant-technical-evaluations | external | www.anthropic.com | Three generations of a take-home evaluation, two of them defeated by successive model releases, and the design properties that survived: multi-hour horizons, wide scoring distributions, and no dependence on a single insight. Relevant to any loop whose gate is a human-graded task. | Three generations of a take-home evaluation, two of them defeated by successive model releases, and the design properties that survived: multi-hour horizons, wide scoring distributions, and no dependence on a single insight. Relevant to any loop whose gate is a human-graded task. | Three generations of a take-home evaluation, two of them defeated by successive model releases, and the design properties that survived: multi-hour horizons, wide scoring distributions, and no dependence on a single insight. Relevant to any loop whose gate is a human-graded task. | Evaluation data is used as the feedback signal for improving loop behavior. Three generations of a take-home evaluation, two of them defeated by successive model releases, and the design properties that survived: multi-hour horizons, wide scoring distributions, and no dependence on a single insight. Relevant to any loo... | Use Designing AI-Resistant Technical Evaluations to measure progress and gate completion with repeatable evidence. | Contextual source from www.anthropic.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,577 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1577 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;escalation | evaluator | evaluation | enabling | practitioner-analysis | B | ok | https://www.anthropic.com/engineering/AI-resistant-technical-evaluations | Designing AI resistant technical evaluations \ Anthropic | What we learned from three iterations of a performance engineering take-home that Claude keeps beating. | Tristan Hume | 2026-01-21 | 2026 | Anthropic Engineering | Anthropic | Author and date read from the post's own byline and dateline on anthropic.com/engineering; kept in code because automatic resolution falls back to the domain for this host. | publisher-meta | 2026-09-26T17:25:52 | |||||||||
ale-0868 | Benchmarks And Evaluation | benchmarks-and-evaluation | Docs | 📚 | Demystifying Evals for AI Agents | https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents | external | www.anthropic.com | Practical guide to combining code-based, model-based, and human graders, with an eight-step roadmap that starts from 20 to 50 tasks drawn from failures already observed in production. Treats the eval suite as a living artifact needing an owner, not a one-time deliverable. | Practical guide to combining code-based, model-based, and human graders, with an eight-step roadmap that starts from 20 to 50 tasks drawn from failures already observed in production. Treats the eval suite as a living artifact needing an owner, not a one-time deliverable. | Practical guide to combining code-based, model-based, and human graders, with an eight-step roadmap that starts from 20 to 50 tasks drawn from failures already observed in production. Treats the eval suite as a living artifact needing an owner, not a one-time deliverable. | Evaluation data is used as the feedback signal for improving loop behavior. Practical guide to combining code-based, model-based, and human graders, with an eight-step roadmap that starts from 20 to 50 tasks drawn from failures already observed in production. Treats the eval suite as a living artifact needing an owner,... | Use Demystifying Evals for AI Agents to measure progress and gate completion with repeatable evidence. | Primary documentation from a platform, SDK, standard, or framework; strong implementation signal. | high | README.md | 1,578 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1578 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;escalation | builder;evaluator | evaluation | enabling | technical-documentation | A | ok | https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents | Demystifying evals for AI agents \ Anthropic | Demystifying evals for AI agents | Mikaela Grace; Jeremy Hadfield; Rodrigo Olivares; Jiri De Jonghe | 2026-01-09 | 2026 | Anthropic Engineering | Anthropic | Author and date read from the post's own byline and dateline on anthropic.com/engineering; kept in code because automatic resolution falls back to the domain for this host. | publisher-meta | 2026-09-26T17:25:52 | |||||||||
ale-0869 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Harness-Bench: Measuring Harness Effects Across Models in Realistic Agent Workflows | https://arxiv.org/abs/2605.27922 | external | arxiv.org | Isolates the variable most benchmarks either hold fixed or abstract away. Runs 106 sandboxed offline tasks across harness configurations and model backends under shared task environments, budgets, and evaluation protocols, while letting each harness keep its native execution behaviour. | Isolates the variable most benchmarks either hold fixed or abstract away. Runs 106 sandboxed offline tasks across harness configurations and model backends under shared task environments, budgets, and evaluation protocols, while letting each harness keep its native execution behaviour. | Isolates the variable most benchmarks either hold fixed or abstract away. Runs 106 sandboxed offline tasks across harness configurations and model backends under shared task environments, budgets, and evaluation protocols, while letting each harness keep its native execution behaviour. | Evaluation data is used as the feedback signal for improving loop behavior. Isolates the variable most benchmarks either hold fixed or abstract away. Runs 106 sandboxed offline tasks across harness configurations and model backends under shared task environments, budgets, and evaluation protocols, while letting each ha... | Use Harness-Bench: Measuring Harness Effects Across Models in Realistic Agent Workflows to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,579 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1579 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2605.27922 | [2605.27922] Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows | LLM agents are increasingly deployed as executable systems that use tools, modify workspaces, and produce concrete artifacts. In such workflows, performance depends not only on the base model, but also on the harness: the system layer that manages context, tools, state, constraints, permissions, tracing, and recovery. ... | Yilun Yao; Xinyu Tan; Chao-Hsuan Liu; Yaoming Li; Zhengyang Wang; Wenhan Yu; Zhewen Tan; Yuxuan Tian; Guangxiang Zhao; Lin Sun; Xiangzheng Zhang; Tong Yang | 2026-05-27 | 2026 | arXiv | arXiv | 16 pages, 4 figures, 11 tables. The first three authors contributed equally | cs.AI | arxiv-api | 2605.27922 | 2026-09-26T17:25:52 | |||||||
ale-0870 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications | https://arxiv.org/abs/2602.22769 | external | arxiv.org | Memory benchmark built on agent-environment trajectories of states, actions, observations, and tool output rather than dialogue transcripts, with synthetic trajectories that scale to arbitrary horizons. Finds existing memory systems lean on lossy similarity search and lose causal structure. | Memory benchmark built on agent-environment trajectories of states, actions, observations, and tool output rather than dialogue transcripts, with synthetic trajectories that scale to arbitrary horizons. Finds existing memory systems lean on lossy similarity search and lose causal structure. | Memory benchmark built on agent-environment trajectories of states, actions, observations, and tool output rather than dialogue transcripts, with synthetic trajectories that scale to arbitrary horizons. Finds existing memory systems lean on lossy similarity search and lose causal structure. | The work turns loop quality into a measurable task or score. Memory benchmark built on agent-environment trajectories of states, actions, observations, and tool output rather than dialogue transcripts, with synthetic trajectories that scale to arbitrary horizons. Finds existing memory systems lean on lossy similarity s... | Use AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,580 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1580 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2602.22769 | [2602.22769] AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications | Large Language Models (LLMs) are increasingly used as autonomous agents in complex, long-horizon applications, where effective memory is critical for sustained performance. Yet existing memory benchmarks are largely dialogue-centric, while real agent memory consists of continuous agent-environment interaction trajector... | Yujie Zhao; Boqin Yuan; Junbo Huang; Haocheng Yuan; Zhongming Yu; Haozhou Xu; Lanxiang Hu; Abhilash Shankarampeta; Zimeng Huang; Wentao Ni; Yuandong Tian; Jishen Zhao | 2026-02-26 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2602.22769 | 2026-09-26T17:25:52 | ||||||||
ale-0871 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents | https://arxiv.org/abs/2604.20006 | external | arxiv.org | Long-term memory benchmark over weeks-to-months histories covering remembering, reasoning, and recommending, plus Forgetting-Aware Memory Accuracy, a metric that penalizes answers built on obsolete or invalidated memory instead of only rewarding successful recall. | Long-term memory benchmark over weeks-to-months histories covering remembering, reasoning, and recommending, plus Forgetting-Aware Memory Accuracy, a metric that penalizes answers built on obsolete or invalidated memory instead of only rewarding successful recall. | Long-term memory benchmark over weeks-to-months histories covering remembering, reasoning, and recommending, plus Forgetting-Aware Memory Accuracy, a metric that penalizes answers built on obsolete or invalidated memory instead of only rewarding successful recall. | The work turns loop quality into a measurable task or score. Long-term memory benchmark over weeks-to-months histories covering remembering, reasoning, and recommending, plus Forgetting-Aware Memory Accuracy, a metric that penalizes answers built on obsolete or invalidated memory instead of only rewarding successful re... | Use From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,581 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1581 | 2026-08-10 | Verify | verify | Gate progress with tests, evals, and evidence. | context;verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://doi.org/10.18653/v1/2026.findings-acl.1337 | [2604.20006] From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents | Personalized agents that interact with users over long periods must maintain persistent memory across sessions and update it as circumstances change. However, existing benchmarks predominantly frame long-term memory evaluation as fact retrieval from past conversations, providing limited insight into agents' ability to ... | Md Nayem Uddin; Kumar Shubham; Eduardo Blanco; Chitta Baral; Gengyu Wang | 2026 | 2026 | Findings of the Association for Computational Linguistics: ACL 2026 | Association for Computational Linguistics | 10.18653/v1/2026.findings-acl.1337 | Published in Findings of the Association for Computational Linguistics: ACL 2026; the linked arXiv record remains available for open access. | cs.CL | Crossref API + DOI record | 2604.20006 | 2026-09-26T17:25:52 | ||||||
ale-0872 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | AlphaEval: Evaluating Agents in Production | https://arxiv.org/abs/2604.12162 | external | arxiv.org | 94 tasks sourced from seven companies running agents in their core business, built against the ways production diverges from benchmarks: implicit requirements, heterogeneous multi-modal inputs with facts split across sources, undeclared domain expertise, long-horizon deliverables, and expert standards that shift over t... | 94 tasks sourced from seven companies running agents in their core business, built against the ways production diverges from benchmarks: implicit requirements, heterogeneous multi-modal inputs with facts split across sources, undeclared domain expertise, long-horizon deliverables, and expert standards that shift over t... | 94 tasks sourced from seven companies running agents in their core business, built against the ways production diverges from benchmarks: implicit requirements, heterogeneous multi-modal inputs with facts split across sources, undeclared domain expertise, long-horizon deliverables, and expert standards that shift over t... | The work turns loop quality into a measurable task or score. 94 tasks sourced from seven companies running agents in their core business, built against the ways production diverges from benchmarks: implicit requirements, heterogeneous multi-modal inputs with facts split across sources, undeclared domain expertise, long... | Use AlphaEval: Evaluating Agents in Production to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,582 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1582 | 2026-08-13 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2604.12162 | [2604.12162] AlphaEval: Evaluating Agents in Production | The rapid deployment of AI agents in commercial settings has outpaced the development of evaluation methodologies that reflect production realities. Existing benchmarks measure agent capabilities through retrospectively curated tasks with well-specified requirements and deterministic metrics -- conditions that diverge ... | Pengrui Lu; Bingyu Xu; Wenjun Zhang; Shengjia Hua; Xuanjian Gao; Ranxiang Ge; Lyumanshan Ye; Linxuan Wu; Yiran Li; Junfei Fish Yu; Yibo Zhang; Ruixin Li; Manxiang Li; Xiao Han; Xiaocong Zhou; Guangyao Chi; Zisheng Chen; Kaishen Chen; Kun Wang; Qihua Xu; Fengyue Meng; Yuchen Ni; Jiajun Li; Jinxiu Liu; Danfeng Zhang; Jin... | 2026-04-14 | 2026 | arXiv | arXiv | cs.CL | arxiv-api | 2604.12162 | 2026-09-26T17:25:52 | ||||||||
ale-0873 | Benchmarks And Evaluation | benchmarks-and-evaluation | Blog | 📝 | A Methodical Approach to Agent Evaluation | https://cloud.google.com/blog/topics/developers-practitioners/a-methodical-approach-to-agent-evaluation | external | cloud.google.com | Splits evaluation into final quality, trajectory, and trust, on the point that an agent can reach a correct answer through a flawed process, a silent failure that outcome-only scoring never surfaces. Pairs human, LLM-judge, code-based, and adversarial methods, then wires the suite into CI as a release gate. | Splits evaluation into final quality, trajectory, and trust, on the point that an agent can reach a correct answer through a flawed process, a silent failure that outcome-only scoring never surfaces. Pairs human, LLM-judge, code-based, and adversarial methods, then wires the suite into CI as a release gate. | Splits evaluation into final quality, trajectory, and trust, on the point that an agent can reach a correct answer through a flawed process, a silent failure that outcome-only scoring never surfaces. Pairs human, LLM-judge, code-based, and adversarial methods, then wires the suite into CI as a release gate. | Evaluation data is used as the feedback signal for improving loop behavior. Splits evaluation into final quality, trajectory, and trust, on the point that an agent can reach a correct answer through a flawed process, a silent failure that outcome-only scoring never surfaces. Pairs human, LLM-judge, code-based, and adve... | Use A Methodical Approach to Agent Evaluation to measure progress and gate completion with repeatable evidence. | Contextual source from cloud.google.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,583 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1583 | 2026-08-13 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;escalation | evaluator | evaluation | enabling | practitioner-analysis | B | ok | https://cloud.google.com/blog/topics/developers-practitioners/a-methodical-approach-to-agent-evaluation | A methodical approach to agent evaluation | Google Cloud Blog | Learn this structured framework to help you build a robust, tailored agent evaluation strategy so you can trust that your agent can move from a proof-of-concept (POC) to production. | Google Cloud Blog | html-meta | 2026-09-26T17:25:52 | ||||||||||||||
ale-0874 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | LoopsBench: From Harness Engineering to Loop Engineering in Coding Agent Evaluation | https://arxiv.org/abs/2608.00267 | external | arxiv.org | Benchmarks sustained execution rather than end state: 112 tasks across 8 languages and 9 domains, each a dependency DAG over separately testable units with source-evidenced prerequisite edges. The runtime releases tests along the ready frontier and keeps completed nodes as standing regression obligations. | Benchmarks sustained execution rather than end state: 112 tasks across 8 languages and 9 domains, each a dependency DAG over separately testable units with source-evidenced prerequisite edges. The runtime releases tests along the ready frontier and keeps completed nodes as standing regression obligations. | Benchmarks sustained execution rather than end state: 112 tasks across 8 languages and 9 domains, each a dependency DAG over separately testable units with source-evidenced prerequisite edges. The runtime releases tests along the ready frontier and keeps completed nodes as standing regression obligations. | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Benchmarks sustained execution rather than end state: 112 tasks across 8 languages and 9 domains, each a dependency DAG over separately testable units with source-evidenced prerequisite edges. The runtime releases tests along the rea... | Use LoopsBench: From Harness Engineering to Loop Engineering in Coding Agent Evaluation to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,584 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1584 | 2026-08-17 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2608.00267 | [2608.00267] LoopsBench: From Harness Engineering to Loop Engineering in Coding Agent Evaluation | Coding agent infrastructure is shifting from harness engineering toward loop engineering as coding agents are deployed for sustained long-horizon software development. Existing benchmarks often center on localized tasks or end-state outcomes, offering limited insight into sustained execution. We introduce LOOPSBENCH, a... | Han Li; Zhemin Fang; Rili Feng; Yingqi Zhao; Jiaheng Liu; Pengfei Gao; He Ye; Dayi Lin; Qingwei Lin; Saravan Rajmohan; Dongmei Zhang | 2026-07-31 | 2026 | arXiv | arXiv | Project page: https://loopsbench.ai/ | cs.SE | arxiv-api | 2608.00267 | 2026-09-26T17:25:52 | |||||||
ale-0875 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Do Codex Skills Save Tokens? Six Controlled GPT-5.6-sol Runs | https://github.com/Phelan164/codex-howto/blob/main/resources/articles/do-codex-skills-save-tokens.md | external | github.com | Asks whether a repository workflow skill earns its context cost, and finds the answer reverses by task size: the no-skill control was cheapest on a small specified backend fix, while a lean engineering-loop skill spent 54% fewer reported tokens on a larger build. Two task classes and six runs, with the measurements pub... | Asks whether a repository workflow skill earns its context cost, and finds the answer reverses by task size: the no-skill control was cheapest on a small specified backend fix, while a lean engineering-loop skill spent 54% fewer reported tokens on a larger build. Two task classes and six runs, with the measurements pub... | Asks whether a repository workflow skill earns its context cost, and finds the answer reverses by task size: the no-skill control was cheapest on a small specified backend fix, while a lean engineering-loop skill spent 54% fewer reported tokens on a larger build. Two task classes and six runs, with the measurements pub... | Context is managed as durable loop state rather than a single prompt payload. Asks whether a repository workflow skill earns its context cost, and finds the answer reverses by task size: the no-skill control was cheapest on a small specified backend fix, while a lean engineering-loop skill spent 54% fewer reported toke... | Use Do Codex Skills Save Tokens? Six Controlled GPT-5.6-sol Runs to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,585 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1585 | 2026-08-20 | Verify | verify | Gate progress with tests, evals, and evidence. | context;budget | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://github.com/Phelan164/codex-howto/blob/main/resources/articles/do-codex-skills-save-tokens.md | codex-howto/resources/articles/do-codex-skills-save-tokens.md at main · Phelan164/codex-howto · GitHub | Engineering-first OpenAI Codex guide: 14 modules, 9 installable skills, reproducible benchmarks, testing, review, orchestration, and a Living Wiki. - codex-howto/resources/articles/do-codex-skills-save-tokens.md at main · Phelan164/codex-howto | Phelan164/codex-howto | GitHub | html-meta | Phelan164/codex-howto | 2026-09-26T17:25:52 | ||||||||||||
ale-0876 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | PhoneHarness: Harnessing Phone-Use Agents through Mixed GUI, CLI, and Tool Actions | https://arxiv.org/abs/2606.14832 | external | arxiv.org | Scores whether the intended side effect actually occurred rather than whether the final screen looks right, which is how most mobile-agent work grades a run. Mixes GUI, device-side commands, and structured tools, so choosing which surface to act through becomes part of the task. | Scores whether the intended side effect actually occurred rather than whether the final screen looks right, which is how most mobile-agent work grades a run. Mixes GUI, device-side commands, and structured tools, so choosing which surface to act through becomes part of the task. | Scores whether the intended side effect actually occurred rather than whether the final screen looks right, which is how most mobile-agent work grades a run. Mixes GUI, device-side commands, and structured tools, so choosing which surface to act through becomes part of the task. | Links loop design to measurable tasks where progress and failure can be compared. Scores whether the intended side effect actually occurred rather than whether the final screen looks right, which is how most mobile-agent work grades a run. Mixes GUI, device-side commands, and structured tools, so choosing which surface... | Use PhoneHarness: Harnessing Phone-Use Agents through Mixed GUI, CLI, and Tool Actions to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,586 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1586 | 2026-08-28 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2606.14832 | [2606.14832] PhoneHarness: Harnessing Phone-Use Agents through Mixed GUI, CLI, and Tool Actions | Phone agents are increasingly expected to complete real mobile workflows rather than merely predict the next screen action. However, much of the current mobile-agent literature still evaluates agents primarily as GUI controllers that observe a screen, emit taps and swipes, and are scored by target app state. Real phone... | Chenxin Li; Zhengyao Fang; Zhengyang Tang; Pengyuan Lyu; Xingran Zhou; Xin Lai; Fei Tang; Liang Wu; Yiduo Guo; Weinong Wang; Junyi Li; Yi Zhang; Yang Ding; Huawen Shen; Sunqi Fan; Shangpin Peng; Zheng Ruan; Anran Zhang; Benyou Wang; Chengquan Zhang; Han Hu | 2026-06-12 | 2026 | arXiv | arXiv | Project Page: https://phoneharness.github.io/ | cs.CL | arxiv-api | 2606.14832 | 2026-09-26T17:25:52 | |||||||
ale-0877 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering | https://arxiv.org/abs/2608.28281 | external | arxiv.org | Benchmarks the loop rather than the agent inside it. A loop can trust a stale progress note, skip needed verification, spend budget in the wrong direction, or stop before the task is safe to submit, and one end-to-end outcome cannot show whether the loop's guidance or the agent's ability decided the result. | Benchmarks the loop rather than the agent inside it. A loop can trust a stale progress note, skip needed verification, spend budget in the wrong direction, or stop before the task is safe to submit, and one end-to-end outcome cannot show whether the loop's guidance or the agent's ability decided the result. | Benchmarks the loop rather than the agent inside it. A loop can trust a stale progress note, skip needed verification, spend budget in the wrong direction, or stop before the task is safe to submit, and one end-to-end outcome cannot show whether the loop's guidance or the agent's ability decided the result. | Verification is promoted from a final check to a loop-control signal. Benchmarks the loop rather than the agent inside it. A loop can trust a stale progress note, skip needed verification, spend budget in the wrong direction, or stop before the task is safe to submit, and one end-to-end outcome cannot show whether the ... | Use LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,587 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1587 | 2026-09-17 | Verify | verify | Gate progress with tests, evals, and evidence. | verification;budget;exit | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2608.28281 | [2608.28281] LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering | Loop Engineering is emerging as a practice for organizing development work around coding agents. Instead of writing each prompt by hand, practitioners design loops that monitor progress, assign work, run checks, and decide what the agent should do next. Even with a capable coding agent, a loop may trust a stale progres... | Yi Wang; Haopeng Zhang; Chengxiang Huang; Rui Dai; Kaikui Liu; Piotr Koniusz; Xiangxiang Chu | 2026-08-28 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2608.28281 | 2026-09-26T17:25:52 | ||||||||
ale-0878 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | BAITBENCH: Measuring Agent Reward Hacking with Optional Shortcuts Planted in ML Tasks | https://arxiv.org/abs/2608.30724 | external | arxiv.org | Plants an optional shortcut in three ML tasks that inflates the public score but fails a hidden test set, so exploiting it breaks no stated rule. 57.1% of runs across seven frontier agents took it, five of the seven above half, and the rate stayed above 50% even when agents were told not to. | Plants an optional shortcut in three ML tasks that inflates the public score but fails a hidden test set, so exploiting it breaks no stated rule. 57.1% of runs across seven frontier agents took it, five of the seven above half, and the rate stayed above 50% even when agents were told not to. | Plants an optional shortcut in three ML tasks that inflates the public score but fails a hidden test set, so exploiting it breaks no stated rule. 57.1% of runs across seven frontier agents took it, five of the seven above half, and the rate stayed above 50% even when agents were told not to. | Links loop design to measurable tasks where progress and failure can be compared. Plants an optional shortcut in three ML tasks that inflates the public score but fails a hidden test set, so exploiting it breaks no stated rule. 57.1% of runs across seven frontier agents took it, five of the seven above half, and the ra... | Use BAITBENCH: Measuring Agent Reward Hacking with Optional Shortcuts Planted in ML Tasks to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,588 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1588 | 2026-09-17 | Verify | verify | Gate progress with tests, evals, and evidence. | verification | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2608.30724 | [2608.30724] BAITBENCH: Measuring Agent Reward Hacking with Optional Shortcuts Planted in ML Tasks | LLM agents are increasingly used to run autonomous ML experiments, iterating on target metrics with little human oversight. Prior work has documented reward hacking in these environments, bringing into question the validity of produced research and the broader safety case for AI R&D. Existing benchmarks do not measure ... | Pradyumna Shyama Prasad; Meiri Anto; Leon Eshuijs; Julian Moncarz; Kaustubh Kislay; Juan J. Vazquez | 2026-08-31 | 2026 | arXiv | arXiv | cs.LG | arxiv-api | 2608.30724 | 2026-09-26T17:25:52 | ||||||||
ale-0879 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents | https://arxiv.org/abs/2609.16305 | external | arxiv.org | Scores whether an agent acts, refuses, or stays calibrated as a long interaction evolves, rather than collapsing behaviour into task or attack success. Evaluates whole user-agent-environment trajectories with adaptive adversarial interaction and stateful tool execution across 22 attack families and 35 scenarios. | Scores whether an agent acts, refuses, or stays calibrated as a long interaction evolves, rather than collapsing behaviour into task or attack success. Evaluates whole user-agent-environment trajectories with adaptive adversarial interaction and stateful tool execution across 22 attack families and 35 scenarios. | Scores whether an agent acts, refuses, or stays calibrated as a long interaction evolves, rather than collapsing behaviour into task or attack success. Evaluates whole user-agent-environment trajectories with adaptive adversarial interaction and stateful tool execution across 22 attack families and 35 scenarios. | The work turns loop quality into a measurable task or score. Scores whether an agent acts, refuses, or stays calibrated as a long interaction evolves, rather than collapsing behaviour into task or attack success. Evaluates whole user-agent-environment trajectories with adaptive adversarial interaction and stateful tool... | Use BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,589 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1589 | 2026-09-21 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2609.16305 | [2609.16305] BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents | Large language model (LLM) agents increasingly operate over long-horizon interactions involving tool use, persistent state, evolving authorization, and external environment feedback. In such settings, safety failures may emerge only after multiple turns, yet existing evaluations often reduce agent behavior to task or a... | Sadia Asif; Mohammad Mohammadi Amiri; Momin Abbas; Tejaswini Pedapati; Prasanna Sattigeri | 2026-09-14 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2609.16305 | 2026-09-26T17:25:52 | ||||||||
ale-0880 | Benchmarks And Evaluation | benchmarks-and-evaluation | Benchmark | 🧪 | Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents | https://arxiv.org/abs/2609.24663 | external | arxiv.org | Endpoint scores miss how self-evolving agents change: memories and skills update throughout a stream, so a capability can emerge and later weaken. Freezes evolving artifacts at checkpoints with base model and tools fixed, tracking when each capability appears and whether later updates keep it. | Endpoint scores miss how self-evolving agents change: memories and skills update throughout a stream, so a capability can emerge and later weaken. Freezes evolving artifacts at checkpoints with base model and tools fixed, tracking when each capability appears and whether later updates keep it. | Endpoint scores miss how self-evolving agents change: memories and skills update throughout a stream, so a capability can emerge and later weaken. Freezes evolving artifacts at checkpoints with base model and tools fixed, tracking when each capability appears and whether later updates keep it. | Checkpointed state makes long-running agent work recoverable across failures. Endpoint scores miss how self-evolving agents change: memories and skills update throughout a stream, so a capability can emerge and later weaken. Freezes evolving artifacts at checkpoints with base model and tools fixed, tracking when each c... | Use Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents to measure progress and gate completion with repeatable evidence. | Benchmark or leaderboard source with repeatable tasks or scores; compare systems only after checking setup and scope. | high | README.md | 1,590 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1590 | 2026-09-27 | Verify | verify | Gate progress with tests, evals, and evidence. | workspace;context;verification;state | researcher;evaluator | evaluation | enabling | benchmark | A | ok | https://arxiv.org/abs/2609.24663 | [2609.24663] Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents | Self-evolving agents convert interaction feedback into persistent artifacts, such as memories or skills, which in turn guide subsequent decisions. As these artifacts are iteratively updated throughout an experience stream, the capabilities they support may evolve. Consequently, endpoint performance alone offers an inco... | Hongqiang Lin; Chao Liu; Xiaofan Bai; Xuan Jin; Yuhong Li; Nenggan Zheng; Xipeng Cao | 2026-09-21 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2609.24663 | 2026-09-26T17:25:52 | ||||||||
ale-0881 | Operations Playbooks | operations-playbooks | Blog | 📝 | Agentic Engineering: The Agent Loop | https://junpingyi.com/books/agentic-engineering/agent-loop/ | external | junpingyi.com | Minimal mental model for the loop underlying agent operation. | Minimal mental model for the loop underlying agent operation. | Minimal mental model for the loop underlying agent operation. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Minimal mental model for the loop underlying agent operation. | Use Agentic Engineering: The Agent Loop to bound risk before recurring or unattended execution. | Contextual source from junpingyi.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,598 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1598 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;budget;escalation;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://junpingyi.com/books/agentic-engineering/agent-loop/ | Chapter 1: The Agent Loop — Agentic Engineering: How to Build AI Agents Like Claude Code | Chapter 1: The Agent Loop from Agentic Engineering: How to Build AI Agents Like Claude Code | junpingyi.com | domain-fallback | 2026-09-26T17:25:52 | |||||||||||||||
ale-0882 | Operations Playbooks | operations-playbooks | Blog | 📝 | The agent loop: ReAct, plan-and-execute, reflection | https://www.kunwar.page/chapter/067-the-agent-loop-react-plan-and-execute-reflection | external | www.kunwar.page | Practical walkthrough of the base loop and common variants. | Practical walkthrough of the base loop and common variants. | Practical walkthrough of the base loop and common variants. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Practical walkthrough of the base loop and common variants. | Use The agent loop: ReAct, plan-and-execute, reflection to bound risk before recurring or unattended execution. | Contextual source from www.kunwar.page; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,599 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1599 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;budget;escalation;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://www.kunwar.page/chapter/067-the-agent-loop-react-plan-and-execute-reflection | Chapter 67: The agent loop: ReAct, plan-and-execute, reflection — The Holy Grail Basic agent loop: generate, check for tool calls, execute tools and loop back, or return final answer on no tool call. ReAct interleaves Thought, Action, and Observation triplets; each Thought improves the next Action choice by externalizi... | An agent is a loop of `model.generate()` calls with tool calls in between. The loop is the entire pattern | kunwar.page | domain-fallback | 2026-09-26T17:25:52 | |||||||||||||||
ale-0883 | Operations Playbooks | operations-playbooks | Blog | 📝 | How to Build an Agent | https://ampcode.com/how-to-build-an-agent | external | ampcode.com | Thorsten Ball's demystification of the inner agent loop: a model, a loop, and enough tokens. | Thorsten Ball's demystification of the inner agent loop: a model, a loop, and enough tokens. | Thorsten Ball's demystification of the inner agent loop: a model, a loop, and enough tokens. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Thorsten Ball's demystification of the inner agent loop: a model, a loop, and enough tokens. | Use How to Build an Agent to bound risk before recurring or unattended execution. | Contextual source from ampcode.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,600 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1600 | Govern | govern | Bound permissions, cost, failure, and escalation. | budget | operator;security | operations | direct | practitioner-analysis | B | ok | https://ampcode.com/notes/how-to-build-an-agent | How to Build an Agent - Amp | Thorsten Ball shows how to build a fully functional, code-editing agent in less than 400 lines of Go. | ampcode.com | domain-fallback | 2026-09-26T17:25:52 | |||||||||||||||
ale-0884 | Operations Playbooks | operations-playbooks | Blog | 📝 | Agentic Coding Recommendations | https://lucumr.pocoo.org/2025/6/12/agentic-coding/ | external | lucumr.pocoo.org | Armin Ronacher's field notes on which practices hold up when agents do most of the work. | Armin Ronacher's field notes on which practices hold up when agents do most of the work. | Armin Ronacher's field notes on which practices hold up when agents do most of the work. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Armin Ronacher's field notes on which practices hold up when agents do most of the work. | Use Agentic Coding Recommendations to bound risk before recurring or unattended execution. | Contextual source from lucumr.pocoo.org; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,601 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1601 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;budget;escalation;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://lucumr.pocoo.org/2025/6/12/agentic-coding/ | Agentic Coding Recommendations | Armin Ronacher's Thoughts and Writings | Current recommendations of agentic coding. | 2025-06-12 | 2025 | Armin Ronacher's Thoughts and Writings | html-meta | 2026-09-26T17:25:52 | |||||||||||||
ale-0885 | Operations Playbooks | operations-playbooks | Blog | 📝 | Coding Agents 101: The Art of Actually Getting Things Done | https://devin.ai/agents101 | external | devin.ai | Practical delegation guidance from the Devin team on scoping tasks agents can actually finish. | Practical delegation guidance from the Devin team on scoping tasks agents can actually finish. | Practical delegation guidance from the Devin team on scoping tasks agents can actually finish. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Practical delegation guidance from the Devin team on scoping tasks agents can actually finish. | Use Coding Agents 101: The Art of Actually Getting Things Done to bound risk before recurring or unattended execution. | Contextual source from devin.ai; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,602 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1602 | Govern | govern | Bound permissions, cost, failure, and escalation. | delegation;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://devin.ai/agents101 | Coding Agents 101: The Art of Actually Getting Things Done | Lessons from engineers and customers on finding success faster with coding agents. | Devin | html-meta | 2026-09-26T17:25:52 | |||||||||||||||
ale-0886 | Operations Playbooks | operations-playbooks | Blog | 📝 | How Anthropic teams use Claude Code | https://claude.com/blog/how-anthropic-teams-use-claude-code | external | claude.com | Cross-team field report of real recurring agent workflows in engineering, security, and data science. | Cross-team field report of real recurring agent workflows in engineering, security, and data science. | Cross-team field report of real recurring agent workflows in engineering, security, and data science. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Cross-team field report of real recurring agent workflows in engineering, security, and data science. | Use How Anthropic teams use Claude Code to bound risk before recurring or unattended execution. | Contextual source from claude.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,603 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1603 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;budget;escalation;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://claude.com/blog/how-anthropic-teams-use-claude-code | How Anthropic teams use Claude Code | Claude by Anthropic | Teams across Anthropic use Claude Code for everything from debugging production issues and navigating unfamiliar codebases to building custom automation tools. Here's how. | Claude | html-meta | 2026-09-26T17:25:52 | |||||||||||||||
ale-0887 | Operations Playbooks | operations-playbooks | Blog | 📝 | How Boris Uses Claude Code | https://howborisusesclaudecode.com/ | external | howborisusesclaudecode.com | Unofficial but concrete compilation of Boris Cherny's autonomous setups: parallel worktrees, auto mode, `/loop`, `/schedule`, dynamic workflows, and `/goal` completion conditions. | Unofficial but concrete compilation of Boris Cherny's autonomous setups: parallel worktrees, auto mode, `/loop`, `/schedule`, dynamic workflows, and `/goal` completion conditions. | Unofficial but concrete compilation of Boris Cherny's autonomous setups: parallel worktrees, auto mode, `/loop`, `/schedule`, dynamic workflows, and `/goal` completion conditions. | Workspace isolation is part of the loop design, not an afterthought. Unofficial but concrete compilation of Boris Cherny's autonomous setups: parallel worktrees, auto mode, `/loop`, `/schedule`, dynamic workflows, and `/goal` completion conditions. | Use How Boris Uses Claude Code to bound risk before recurring or unattended execution. | Contextual source from howborisusesclaudecode.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,604 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1604 | Govern | govern | Bound permissions, cost, failure, and escalation. | objective;trigger;workspace;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://howborisusesclaudecode.com/ | Boris Cherny's Claude Code Tips — How He Actually Uses It (136+ Tips) | 136+ tips from Boris Cherny, creator of Claude Code, on his daily workflow: CLAUDE.md, worktrees, plan mode, hooks, subagents, and more. | @CarolinaCherry | How Boris Uses Claude Code | html-meta | 2026-09-26T17:25:52 | ||||||||||||||
ale-0888 | Operations Playbooks | operations-playbooks | Blog | 📝 | Agent of the Day: Copilot Agent PR Analysis | https://github.github.com/gh-aw/blog/2026-05-26-agent-of-the-day/ | external | github.github.com | Official walkthrough of a daily scheduled agentic workflow that ingests PR data, analyzes it, and publishes findings to a Discussion, a concrete recurring loop with trigger, intake, analysis, and output. | Official walkthrough of a daily scheduled agentic workflow that ingests PR data, analyzes it, and publishes findings to a Discussion, a concrete recurring loop with trigger, intake, analysis, and output. | Official walkthrough of a daily scheduled agentic workflow that ingests PR data, analyzes it, and publishes findings to a Discussion, a concrete recurring loop with trigger, intake, analysis, and output. | Primary-source operational guidance rather than commentary. Official walkthrough of a daily scheduled agentic workflow that ingests PR data, analyzes it, and publishes findings to a Discussion, a concrete recurring loop with trigger, intake, analysis, and output. | Use Agent of the Day: Copilot Agent PR Analysis to bound risk before recurring or unattended execution. | Contextual source from github.github.com; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,605 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1605 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake | operator;security | operations | direct | practitioner-analysis | B | ok | https://github.github.com/gh-aw/blog/2026-05-26-agent-of-the-day/ | Agent of the Day – May 26, 2026 | GitHub Agentic Workflows | Copilot Agent PR Analysis: a daily workflow that monitors GitHub Copilot coding agent performance across pull requests | 2026 | GitHub Agentic Workflows | html-meta | 2026-09-26T17:25:52 | ||||||||||||||
ale-0889 | Operations Playbooks | operations-playbooks | Paper | 📄 | Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows | https://arxiv.org/abs/2607.07052 | external | arxiv.org | Production lifecycle in which repeatedly validated agent-loop behaviors are promoted into deterministic workflows and demoted on regression, cutting per-incident agent cost by over 70% across eight months of a cloud AIOps system. | Production lifecycle in which repeatedly validated agent-loop behaviors are promoted into deterministic workflows and demoted on regression, cutting per-incident agent cost by over 70% across eight months of a cloud AIOps system. | Production lifecycle in which repeatedly validated agent-loop behaviors are promoted into deterministic workflows and demoted on regression, cutting per-incident agent cost by over 70% across eight months of a cloud AIOps system. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Production lifecycle in which repeatedly validated agent-loop behaviors are promoted into deterministic workflows and demoted on regression, cutting per-incident agent cost by over 70% across eight months of a cloud AIOps system. | Use Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows to bound risk before recurring or unattended execution. | Research source arXiv:2607.07052; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,606 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1606 | Govern | govern | Bound permissions, cost, failure, and escalation. | budget | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.07052 | [2607.07052] Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows in Production | AI agents deployed for IT operations are typically permanent cost centers because every execution requires full LLM inference, even for previously solved problems. This paper introduces progressive crystallization, a lifecycle that treats agent exploration as a discovery mechanism rather than a permanent execution mode... | Arun Malik | 2026-07-08 | 2026 | arXiv | arXiv | Conference-style paper; 10 pages (estimated from manuscript formatting if applicable); focuses on agentic AI, AIOps, workflow automation, deterministic execution, and LLM cost optimization | cs.SE | arxiv-api | 2607.07052 | 2026-09-26T17:25:52 | ||||||||
ale-0890 | Operations Playbooks | operations-playbooks | Paper | 📄 | Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems | https://arxiv.org/abs/2607.08010 | external | arxiv.org | Production tool-making pipeline that mines live execution traces to compile recurring SOP steps into validated, versioned tools agents call instead of regenerating code, cutting median latency 42% and errors up to 53% in a fulfillment-center alarm-triage deployment. | Production tool-making pipeline that mines live execution traces to compile recurring SOP steps into validated, versioned tools agents call instead of regenerating code, cutting median latency 42% and errors up to 53% in a fulfillment-center alarm-triage deployment. | Production tool-making pipeline that mines live execution traces to compile recurring SOP steps into validated, versioned tools agents call instead of regenerating code, cutting median latency 42% and errors up to 53% in a fulfillment-center alarm-triage deployment. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Production tool-making pipeline that mines live execution traces to compile recurring SOP steps into validated, versioned tools agents call instead of regenerating code, cutting median latency 42% and errors up to 53% in a fulfillme... | Use Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems to bound risk before recurring or unattended execution. | Research source arXiv:2607.08010; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,607 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1607 | Govern | govern | Bound permissions, cost, failure, and escalation. | intake;workspace | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.08010 | [2607.08010] Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems | Production LLM agents often waste latency and reliability by regenerating code for the same procedural steps on every request. We replace this inference-time coding loop with an agentic tool-making pipeline that compiles repeated SOP steps into validated, versioned tools before deployment. The tool-maker grounds synthe... | Kalle Kujanpää; Ning Liu; Shahnawaz Alam; Yeshwanth Reddy Sura; Tianyu Yang; Kristina Klinkner; Shervin Malmasi | 2026-07-09 | 2026 | arXiv | arXiv | To appear at EMNLP 2026 Industry Track | cs.CL | arxiv-api | 2607.08010 | 2026-09-26T17:25:52 | ||||||||
ale-0891 | Operations Playbooks | operations-playbooks | Blog | 📝 | AI Loop Engineering: Build Autonomous Agents with Claude Code /goal and Routines | https://www.sabrina.dev/p/loop-engineering-claude-code-goal-routines | external | www.sabrina.dev | Sabrina Ramonov's practitioner walkthrough of building verified loops with Claude Code `/goal` and cloud routines: a six-part loop-engineering framework, five worked examples each with a verifiable end state, and a production daily support-ticket cleanup routine whose independent checker agents illustrate why the check... | Sabrina Ramonov's practitioner walkthrough of building verified loops with Claude Code `/goal` and cloud routines: a six-part loop-engineering framework, five worked examples each with a verifiable end state, and a production daily support-ticket cleanup routine whose independent checker agents illustrate why the check... | Sabrina Ramonov's practitioner walkthrough of building verified loops with Claude Code `/goal` and cloud routines: a six-part loop-engineering framework, five worked examples each with a verifiable end state, and a production daily support-ticket cleanup routine whose independent checker agents illustrate why the check... | Verification is promoted from a final check to a loop-control signal. Sabrina Ramonov's practitioner walkthrough of building verified loops with Claude Code `/goal` and cloud routines: a six-part loop-engineering framework, five worked examples each with a verifiable end state, and a production daily support-ticket cle... | Use AI Loop Engineering: Build Autonomous Agents with Claude Code /goal and Routines to bound risk before recurring or unattended execution. | Contextual source from www.sabrina.dev; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,608 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1608 | Govern | govern | Bound permissions, cost, failure, and escalation. | objective;verification;state | operator;security | operations | direct | practitioner-analysis | B | ok | https://www.sabrina.dev/p/loop-engineering-claude-code-goal-routines | AI Loop Engineering: Build Autonomous Agents with Claude Code /goal + Routines | What loop engineering means in 2026, how to use the Claude Code /goal command, and how to build your first autonomous AI agent with a routine. | Sabrina Ramonov 🍄 | sabrina.dev | html-meta | 2026-09-26T17:25:52 | ||||||||||||||
ale-0892 | Operations Playbooks | operations-playbooks | Paper | 📄 | Agent Delivery Engineering Predictive Reliability Framework | https://arxiv.org/abs/2607.07689 | external | arxiv.org | Proactive health-trajectory prediction for long-horizon multi-agent systems, aggregating 20 heterogeneous signals across five layers into a trust-margin metric with 8-hour forecasts at 76.8% direction accuracy, detecting degradation concealed by normal surface metrics across 15 days of production traffic. | Proactive health-trajectory prediction for long-horizon multi-agent systems, aggregating 20 heterogeneous signals across five layers into a trust-margin metric with 8-hour forecasts at 76.8% direction accuracy, detecting degradation concealed by normal surface metrics across 15 days of production traffic. | Proactive health-trajectory prediction for long-horizon multi-agent systems, aggregating 20 heterogeneous signals across five layers into a trust-margin metric with 8-hour forecasts at 76.8% direction accuracy, detecting degradation concealed by normal surface metrics across 15 days of production traffic. | The work separates roles across agents, verifiers, or orchestration layers. Proactive health-trajectory prediction for long-horizon multi-agent systems, aggregating 20 heterogeneous signals across five layers into a trust-margin metric with 8-hour forecasts at 76.8% direction accuracy, detecting degradation concealed b... | Use Agent Delivery Engineering Predictive Reliability Framework to bound risk before recurring or unattended execution. | Research source arXiv:2607.07689; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,609 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1609 | Govern | govern | Bound permissions, cost, failure, and escalation. | delegation | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.07689 | [2607.07689] Agent Delivery Engineering Predictive Reliability Framework | Long-horizon LLM multi-agent systems face reliability risks invisible to infrastructure monitoring. We propose the ADE Predictive Reliability Framework (ADE-PRF), enabling proactive health trajectory prediction from passive degradation detection. ADE-PRF aggregates 20 heterogeneous signals across five layers into a Tru... | Dexing Liu | 2026-07-08 | 2026 | arXiv | arXiv | 117pages,83figures | cs.MA | arxiv-api | 2607.07689 | 2026-09-26T17:25:52 | ||||||||
ale-0893 | Operations Playbooks | operations-playbooks | Paper | 📄 | CADAQUES: A Cost-Aware Dual Architecture for Query-Efficient Autonomous Discovery | https://arxiv.org/abs/2607.16127 | external | arxiv.org | Makes cost part of the loop contract: an Oracle executes queries, a Driver chooses them, a shared vector budget covers wall time, CPU, money, and LLM tokens, and an append-only ledger records declared versus settled cost for every transaction. The Ising-model case study shows how mixed-fidelity scheduling can outperfor... | Makes cost part of the loop contract: an Oracle executes queries, a Driver chooses them, a shared vector budget covers wall time, CPU, money, and LLM tokens, and an append-only ledger records declared versus settled cost for every transaction. The Ising-model case study shows how mixed-fidelity scheduling can outperfor... | Makes cost part of the loop contract: an Oracle executes queries, a Driver chooses them, a shared vector budget covers wall time, CPU, money, and LLM tokens, and an append-only ledger records declared versus settled cost for every transaction. The Ising-model case study shows how mixed-fidelity scheduling can outperfor... | The trigger or cadence is explicit, making the workflow recurring rather than one-off. Makes cost part of the loop contract: an Oracle executes queries, a Driver chooses them, a shared vector budget covers wall time, CPU, money, and LLM tokens, and an append-only ledger records declared versus settled cost for every tr... | Use CADAQUES: A Cost-Aware Dual Architecture for Query-Efficient Autonomous Discovery to bound risk before recurring or unattended execution. | Research source arXiv:2607.16127; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,610 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1610 | 2026-07-20 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;verification;budget | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.16127 | [2607.16127] CADAQUES: A Cost-Aware Dual Architecture for Query-Efficient Autonomous Discovery | Autonomous discovery systems couple a resource that answers queries (a simulator, instrument, or analytic model) to an algorithm that selects what to query next. Most software frameworks for this loop inherit the control structure of numerical optimization: campaigns run for a fixed number of iterations, query costs ar... | Jorge Bravo-Abad | 2026-07-17 | 2026 | arXiv | arXiv | physics.comp-ph | arxiv-api | 2607.16127 | 2026-09-26T17:25:52 | ||||||||
ale-0894 | Operations Playbooks | operations-playbooks | Tool | 🧰 | rocketplaneIO | https://github.com/olemeyer/rocketplaneIO | external | github.com | Self-hosted AI SRE for Kubernetes whose copilot investigates autonomously via eBPF traces, logs, and live service maps, but can only act through named, reversible, risk-graded operations. | Self-hosted AI SRE for Kubernetes whose copilot investigates autonomously via eBPF traces, logs, and live service maps, but can only act through named, reversible, risk-graded operations. | Self-hosted AI SRE for Kubernetes whose copilot investigates autonomously via eBPF traces, logs, and live service maps, but can only act through named, reversible, risk-graded operations. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Self-hosted AI SRE for Kubernetes whose copilot investigates autonomously via eBPF traces, logs, and live service maps, but can only act through named, reversible, risk-graded operations. | Use rocketplaneIO to bound risk before recurring or unattended execution. | Inspectable GitHub source; popularity is context, not proof of reliability. | medium | README.md | 1,611 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1611 | Govern | govern | Bound permissions, cost, failure, and escalation. | trigger;intake;budget;escalation;exit | builder;operator;security | operations | direct | source-implementation | A | ok | https://github.com/olemeyer/rocketplaneIO | GitHub - olemeyer/rocketplaneIO: Self-hosted AI SRE for Kubernetes — zero-instrumentation eBPF observability plus a copilot that fixes issues through guardrailed, self-verifying actions. BYO-LLM, air-gapped capable. · GitHub | Self-hosted AI SRE for Kubernetes — zero-instrumentation eBPF observability plus a copilot that fixes issues through guardrailed, self-verifying actions. BYO-LLM, air-gapped capable. - olemeyer/rocketplaneIO | olemeyer/rocketplaneIO | GitHub | html-meta | olemeyer/rocketplaneIO | 2026-09-26T17:25:52 | |||||||||||||
ale-0895 | Operations Playbooks | operations-playbooks | Blog | 📝 | Migrating a Production AI Agent to GPT-5.6 | https://ploy.ai/blog/migrating-a-production-ai-agent-to-gpt-5-6 | external | ploy.ai | Lorenzo Gentile's July 2026 case study of swapping the model inside Ploy's production website-building agent (plans pages, writes components, screenshots its own work, decides when done): roughly a third of initial cross-model eval failures traced to Opus-specific harness assumptions rather than the new model, GPT-5.6 ... | Lorenzo Gentile's July 2026 case study of swapping the model inside Ploy's production website-building agent (plans pages, writes components, screenshots its own work, decides when done): roughly a third of initial cross-model eval failures traced to Opus-specific harness assumptions rather than the new model, GPT-5.6 ... | Lorenzo Gentile's July 2026 case study of swapping the model inside Ploy's production website-building agent (plans pages, writes components, screenshots its own work, decides when done): roughly a third of initial cross-model eval failures traced to Opus-specific harness assumptions rather than the new model, GPT-5.6 ... | Evaluation data is used as the feedback signal for improving loop behavior. Lorenzo Gentile's July 2026 case study of swapping the model inside Ploy's production website-building agent (plans pages, writes components, screenshots its own work, decides when done): roughly a third of initial cross-model eval failures tra... | Use Migrating a Production AI Agent to GPT-5.6 to bound risk before recurring or unattended execution. | Contextual source from ploy.ai; useful for practice signals or boundary conditions, not independent validation. | contextual | README.md | 1,612 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1612 | Govern | govern | Bound permissions, cost, failure, and escalation. | workspace;verification;state;exit | operator;security | operations | direct | practitioner-analysis | B | ok | https://ploy.ai/blog/migrating-a-production-ai-agent-to-gpt-5-6 | Migrating a production AI agent to GPT-5.6 | Ploy | For four months, no frontier model beat Claude Opus in our production evals. GPT-5.6 did. This is what we learned while migrating. | Ploy | html-meta | 2026-09-26T17:25:52 | |||||||||||||||
ale-0896 | Operations Playbooks | operations-playbooks | Paper | 📄 | Coding-agents can replicate scientific machine learning papers | https://arxiv.org/abs/2607.02134 | external | arxiv.org | Runs coding agents independently 12 times across four scientific machine-learning papers; all workspaces satisfy the study's completion criteria, and 158 implementation targets are linked to report evidence, offering a concrete playbook for evidence-backed research replication. | Runs coding agents independently 12 times across four scientific machine-learning papers; all workspaces satisfy the study's completion criteria, and 158 implementation targets are linked to report evidence, offering a concrete playbook for evidence-backed research replication. | Runs coding agents independently 12 times across four scientific machine-learning papers; all workspaces satisfy the study's completion criteria, and 158 implementation targets are linked to report evidence, offering a concrete playbook for evidence-backed research replication. | Translates agent-loop ideas into operator-facing workflows for repeated delegated work. Runs coding agents independently 12 times across four scientific machine-learning papers; all workspaces satisfy the study's completion criteria, and 158 implementation targets are linked to report evidence, offering a concrete play... | Use Coding-agents can replicate scientific machine learning papers to bound risk before recurring or unattended execution. | Research source arXiv:2607.02134; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,613 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1613 | 2026-07-17 | Govern | govern | Bound permissions, cost, failure, and escalation. | workspace;exit | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.02134 | [2607.02134] Coding-agents can replicate scientific machine learning papers | Scientific machine learning papers typically make computational claims, e.g., that the relative mean square error is less than 5% or that the 95% predictive credible interval covers the test data. A coding agent can be prompted to replicate those claims from paper materials alone, but the prompt does not by itself reli... | Atharva Hans; Ilias Bilionis | 2026-07-02 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.02134 | 2026-09-26T17:25:52 | ||||||||
ale-0897 | Operations Playbooks | operations-playbooks | Paper | 📄 | Agentic ERP: Multi-Agent Large Language Model Architecture for Autonomous Enterprise Resource Planning | https://arxiv.org/abs/2607.17331 | external | arxiv.org | Reference architecture that decomposes ERP operations into role-aligned LLM agents under a graph-based planner-executor-reflector-responder orchestration with a risk-tiered human-in-the-loop harness, evaluated across six orchestration paradigms and a 365-day agent-in-the-loop simulation against rule-based RPA, recurrin... | Reference architecture that decomposes ERP operations into role-aligned LLM agents under a graph-based planner-executor-reflector-responder orchestration with a risk-tiered human-in-the-loop harness, evaluated across six orchestration paradigms and a 365-day agent-in-the-loop simulation against rule-based RPA, recurrin... | Reference architecture that decomposes ERP operations into role-aligned LLM agents under a graph-based planner-executor-reflector-responder orchestration with a risk-tiered human-in-the-loop harness, evaluated across six orchestration paradigms and a 365-day agent-in-the-loop simulation against rule-based RPA, recurrin... | Control flow is represented as an inspectable graph rather than an opaque prompt loop. Reference architecture that decomposes ERP operations into role-aligned LLM agents under a graph-based planner-executor-reflector-responder orchestration with a risk-tiered human-in-the-loop harness, evaluated across six orchestratio... | Use Agentic ERP: Multi-Agent Large Language Model Architecture for Autonomous Enterprise Resource Planning to bound risk before recurring or unattended execution. | Research source arXiv:2607.17331; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,614 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1614 | 2026-07-22 | Govern | govern | Bound permissions, cost, failure, and escalation. | delegation;escalation | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.17331 | [2607.17331] Agentic ERP: Multi-Agent Large Language Model Architecture for Autonomous Enterprise Resource Planning | Enterprise Resource Planning (ERP) systems record transactions reliably but still delegate almost all operational decision-making to human specialists, because classical rule-based automation cannot reason about exceptions and monolithic AI assistants degrade when asked to coordinate across functional boundaries. This ... | Zhihao Liu; Tianyu Wang; Xi Vincent Wang; Lihui Wang | 2026-07-19 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.17331 | 2026-09-26T17:25:52 | ||||||||
ale-0898 | Operations Playbooks | operations-playbooks | Paper | 📄 | The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agents | https://arxiv.org/abs/2607.06906 | external | arxiv.org | Controlled comparison across six foundation models showing the harness, not model choice, dominates agent economics: a redesigned Writer Agent Harness cuts blended cost per task 41%, tokens 38%, and wall-clock 44%, arguing the harness is the one component whose efficiency multiplies across every model an org runs. | Controlled comparison across six foundation models showing the harness, not model choice, dominates agent economics: a redesigned Writer Agent Harness cuts blended cost per task 41%, tokens 38%, and wall-clock 44%, arguing the harness is the one component whose efficiency multiplies across every model an org runs. | Controlled comparison across six foundation models showing the harness, not model choice, dominates agent economics: a redesigned Writer Agent Harness cuts blended cost per task 41%, tokens 38%, and wall-clock 44%, arguing the harness is the one component whose efficiency multiplies across every model an org runs. | Orchestration and control flow are made explicit and inspectable. Controlled comparison across six foundation models showing the harness, not model choice, dominates agent economics: a redesigned Writer Agent Harness cuts blended cost per task 41%, tokens 38%, and wall-clock 44%, arguing the harness is the one componen... | Use The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agents to bound risk before recurring or unattended execution. | Research source arXiv:2607.06906; inspect its method and evaluation before treating results as production evidence. | medium | README.md | 1,615 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1615 | 2026-07-22 | Govern | govern | Bound permissions, cost, failure, and escalation. | delegation;budget | researcher;evaluator;operator;security | operations | direct | research-preprint | A | ok | https://arxiv.org/abs/2607.06906 | [2607.06906] The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI | Agentic AI development today runs on token maxing: buying capability with tokens -- longer reasoning traces, more turns, wider tool payloads, bigger replayed contexts -- so tokens per task grow faster than task value. Falling per-token prices mask the pattern; total spend rises anyway. We argue the decisive lever again... | Muayad Sayed Ali; Aliaksandra Novik; Anji Boddupally; Artem Yavorskyi; Chris Nickerson; Daniel Rica; Emily DuGranrut; Felix Leung; Garrett Prince; Grace Barnett; Heath Robinson; Hosain Al Ahmad; Jesse Resnick; Juan Carlos Farah; Jyothi Swaroop Meruga; Leonid Kuznetsov; Luke Gorham; Marie Schmoll; Michael Paciullo; Saum... | 2026-07-08 | 2026 | arXiv | arXiv | cs.AI | arxiv-api | 2607.06906 | 2026-09-26T17:25:52 | ||||||||
ale-0899 | Operations Playbooks | operations-playbooks | Tool | 🧰 | claude-thermos | https://github.com/izeigerman/claude-thermos | external | github.com | Keeps a Claude Code session's prompt cache warm while the main agent waits on long-running subagents, so an orchestrator that blocks for more than a few minutes does not silently pay to rebuild its cache on every resume. | Keeps a Claude Code session's prompt cache warm while the main agent waits on long-running subagents, so an orchestrator that blocks for more than a few minutes does not silently pay to rebuild its cache on every resume. | Keeps a Claude Code session's prompt cache warm while the main agent waits on long-running subagents, so an orchestrator that blocks for more than a few minutes does not silently pay to rebuild its cache on every resume. | The work separates roles across agents, verifiers, or orchestration layers. Keeps a Claude Code session's prompt cache warm while the main agent waits on long-running subagents, so an orchestrator that blocks for more than a few minutes does not silently pay to rebuild its cache on every resume. | Use claude-thermos to bound risk before recurring or unattended execution. | Inspectable GitHub source; popularity is context, not proof of reliability. | medium | README.md | 1,616 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1616 | 2026-07-24 | Govern | govern | Bound permissions, cost, failure, and escalation. | delegation | builder;operator;security | operations | direct | source-implementation | A | ok | https://github.com/izeigerman/claude-thermos | GitHub - izeigerman/claude-thermos: Keeps your Claude session warm for you · GitHub | Keeps your Claude session warm for you. Contribute to izeigerman/claude-thermos development by creating an account on GitHub. | izeigerman/claude-thermos | GitHub | html-meta | izeigerman/claude-thermos | 2026-09-26T17:25:52 | ||||||||||||
ale-0900 | Operations Playbooks | operations-playbooks | Tool | 🧰 | Worktrunk | https://github.com/max-sixty/worktrunk | external | github.com | CLI for Git worktree management built for running several coding agents in parallel, giving each agent an isolated checkout so concurrent work does not collide in a shared working directory. | CLI for Git worktree management built for running several coding agents in parallel, giving each agent an isolated checkout so concurrent work does not collide in a shared working directory. | CLI for Git worktree management built for running several coding agents in parallel, giving each agent an isolated checkout so concurrent work does not collide in a shared working directory. | Workspace isolation is part of the loop design, not an afterthought. CLI for Git worktree management built for running several coding agents in parallel, giving each agent an isolated checkout so concurrent work does not collide in a shared working directory. | Use Worktrunk to bound risk before recurring or unattended execution. | Inspectable GitHub source; popularity is context, not proof of reliability. | medium | README.md | 1,617 | https://github.com/ChaoYue0307/awesome-loop-engineering/blob/main/README.md#L1617 | 2026-07-28 | Govern | govern | Bound permissions, cost, failure, and escalation. | workspace | builder;operator;security | operations | direct | source-implementation | A | ok | https://github.com/max-sixty/worktrunk | GitHub - max-sixty/worktrunk: Worktrunk is a CLI for Git worktree management, designed for parallel AI agent workflows · GitHub | Worktrunk is a CLI for Git worktree management, designed for parallel AI agent workflows - max-sixty/worktrunk | max-sixty/worktrunk | GitHub | html-meta | max-sixty/worktrunk | 2026-09-26T17:25:52 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.