Pith. sign in

REVIEW 4 major objections 6 minor 24 references

This paper claims that long-horizon agent capability is a property of the whole model–harness system, not the model alone, proved by an audited task-state loop that lifts a single model's WeaveBench pass rate from 51.8% to 80.7%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 17:35 UTC pith:QWIVUWZK

load-bearing objection Real architecture result with matched WeaveBench/Terminal-Bench gains; OSWorld is confounded and the load-bearing auditor accuracy is unmeasured, but it deserves a serious referee. the 4 major comments →

arxiv 2608.01964 v1 pith:QWIVUWZK submitted 2026-08-03 cs.CV

LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

Ziyu Ma , Hailang Huang , Shun Zou , Yong Wang , Shidong Yang , Yiming Hu , Fei Wei , XiangXiang Chu This is my paper
classification cs.CV
keywords long-horizon agentstask-state managementManage-Execute-Audit loopindependent verificationcontext rotcomputer-use agentsagent harnessLLM agents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the main obstacle to long-horizon agent tasks is not any single step but the loss and corruption of task state across many interdependent steps. It proposes LongHorizon-Harness, a system that keeps task state outside the execution context, runs each subtask in a fresh context, and lets a read-only auditor verify the environment before updating the state. On matched comparisons with the same backbone model, this Manage-Execute-Audit loop raises Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench 2.1, and from 2.8% to 8.3% on OSWorld 2.0. The result matters because it suggests that a large share of agent capability can be credited to the harness that organizes, verifies, and accumulates progress, not merely to the underlying model.

Core claim

LongHorizon-Harness reformulates long-horizon execution as task-state management. The manager maintains a structured ledger of requirements, artifacts, and facts; the executor performs a single bounded subtask in a fresh context; and a read-only auditor inspects the resulting environment and returns a report. Only audit-verified findings advance the persistent state, and the executor's raw trajectory is discarded after every round. Across WeaveBench, OSWorld 2.0, and Terminal-Bench 2.1, the same backbone model completes tasks at substantially higher rates when wrapped in this loop, and the gains transfer to another backbone. The paper's central claim is that explicit, externally audited stat

What carries the argument

The Manage-Execute-Audit (MEA) loop is the load-bearing mechanism: a manager owns persistent task state and writes bounded subtask contracts; a fresh-context executor performs one contract; a read-only auditor independently inspects the resulting environment and produces the only evidence allowed to update the state. The task-state ledger itself—records marked completed, pending, blocked, or untrusted with references to audit evidence—is the central object. The loop's design means the executor's interaction history is discarded each round, so only compact, verified state persists across rounds.

Load-bearing premise

The load-bearing premise is that the read-only auditor can reliably certify completion and integrity from environment inspection; the paper reports no measurement of auditor error rates, and because the auditor is usually the same model as the executor, its independence is procedural rather than cognitive—if the audit is wrong, the external state becomes a source of confidently verified wrong answers.

What would settle it

A concrete test: build a task where the executor produces a visually or textually plausible artifact that violates an acceptance criterion, and measure whether the auditor flags it. If auditor false-negative rates are high, or if replacing auditor verdicts with executor self-reports leaves performance unchanged, the audit step is not load-bearing and the gains must come from decomposition or fresh contexts instead.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Agent capability is a property of the model–harness system: a stronger harness can raise task-level performance for a fixed model, and stronger models plus explicit state management give complementary gains (e.g., Qwen with the harness beats Claude Opus with the baseline Claude Code harness on the same WeaveBench games subset).
  • Long-horizon gains appear across GUI+CLI, desktop, and pure command-line domains, so the mechanism is not specific to visual computer use.
  • The harness mainly raises the failure floor: tasks at or near zero baseline completion recover to substantial scores, while already-solved tasks see smaller changes.
  • Token cost is not a fixed multiplier: it can rise (WeaveBench ~2.3x, OSWorld ~3.6x output tokens) or fall (Terminal-Bench 24% fewer tokens) depending on model and task, because manager overhead is small and the auditor is the main added cost.
  • Independent auditing can detect an incorrect result and trigger recovery, but cannot supply a missing primitive capability; the benefit concentrates where the bottleneck is long-horizon reliability rather than a single skill.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the auditor's verdicts are the only gateway into persistent state, then auditor accuracy, not model capability, becomes the effective ceiling on what the harness can certify. The paper does not measure auditor false positives or false negatives, so a direct test would be to inject a known-wrong executor report and see whether the audit catches it.
  • Because manager, executor, and auditor are usually the same backbone model, the 'independence' of the audit is procedural (fresh context, read-only tools, no trajectory) rather than cognitive. Using a different, possibly stronger or more conservative model as auditor is a natural extension the paper leaves implicit.
  • The framework suggests that progress on long-horizon benchmarks may be attainable by reorganizing inference around state verification even without new model training—an implication that could be tested by applying the same MEA wrapper to more backbones and seeing whether the gains persist.
  • The failure analysis points to a next bottleneck: tasks whose correctness depends on hidden thresholds or semantics not recoverable from visible state can still end in confidently verified wrong answers, so verifier design (knowing what state is authoritative) may become the next limiting factor.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes LongHorizon-Harness, a harness that separates long-horizon task execution from task-state management via a Manage-Execute-Audit (MEA) loop. A manager maintains an explicit task-state ledger and proposes one bounded subtask per round; an executor performs the subtask in a fresh context; a read-only auditor inspects the environment and produces an audit report that is the only cross-round memory. The paper reports large improvements over matched baselines for Qwen 3.7-Plus on WeaveBench (51.8→80.7 PassRate), Terminal-Bench 2.1 (69.7→77.2), and an OSWorld 2.0 subset with Claude Opus 4.7 (20.6→35.3), plus a smaller OSWorld 2.0 gain (2.8→8.3) in a hybrid GUI+CLI setting. It also provides domain/tag-level breakdowns and qualitative case studies.

Significance. If the results hold, the paper makes a significant empirical case that the harness—not just the model—is a primary determinant of long-horizon agent performance. The matched WeaveBench and Terminal-Bench comparisons are strong: the same backbone and execution backend differ only by the MEA layer. The per-domain, per-tag, and per-task analyses are useful, and the case studies illustrate the intended failure-recovery mechanisms. The main scientific risk is that the central mechanism—'independently verified facts'—rests on an unmeasured auditor, and the flagship OSWorld comparison confounds tool access with state management. The paper is honest about residual failure modes, but those admissions also underline the load-bearing premise.

major comments (4)
  1. [§2.4, Eq. (3); Appendix B.2.2] The load-bearing premise is that the read-only auditor produces reliable verdicts, since the manager's state update in Eq. (1) consumes only audit reports. No auditor precision/recall is reported, and in all runs the auditor is the same backbone as the executor (§3.1), so 'independent' is procedural, not cognitive. Appendix B.2.2 concedes 'plausible evidence can be mistaken for completed evidence' and 'a misinterpreted contract can still lead to a confidently verified wrong answer.' Please add a quantitative audit-error analysis (e.g., compare audit verdicts against benchmark ground truth on a sample) or an ablation with a different/stronger auditor. Without this, the gains could be attributed to task decomposition and fresh-context execution rather than to verified-state management.
  2. [Table 2; Fig. 1; §A.2] The OSWorld 2.0 comparison is confounded: the baseline is the official single-action GUI setting, while LH-Harness uses a hybrid GUI+CLI tool pool (§A.2). The 2.8→8.3 gain therefore conflates added CLI tool access with the MEA loop. Figure 1's caption even claims improvement 'with the same backbone and execution backend' across all three benchmarks, which is not true for OSWorld. Please add a matched baseline (same hybrid tools without MEA, or a single-action LH-Harness) and correct the caption. This does not undercut the WeaveBench/Terminal-Bench results, but it weakens the OSWorld-specific and Fig. 4 cost-frontier claims.
  3. [§3.1; Table 1; Fig. 3] Evaluation reporting needs tightening. WeaveBench and OSWorld appear to be single runs; please report variance/confidence intervals or explicitly state that the benchmarks are deterministic under the evaluation protocol. Table 1's note says our runs use root privileges while official results use a regular user; please state explicitly whether the Qwen 3.7-Plus Claude Code baseline row (51.8%) was run under the same root privileges and the same judge configuration as LH-Harness, since this is the matched comparison on which the main WeaveBench claim rests. For Terminal-Bench, Fig. 3 should clearly identify which values are externally reported versus this paper's runs.
  4. [§2.2, Eq. (1); §3.3] The mechanism claim would be strengthened by an explicit ablation against a simpler fresh-context decomposition baseline without an audit-based state ledger (e.g., manager + executor only, or a memory-summarization wrapper). The current matched baseline (Claude Code) already includes planning and subagents, so it is not clear that the explicit ledger, rather than bounded contracts, drives the gains. This is a local issue and can be addressed with an additional experiment or a careful discussion.
minor comments (6)
  1. [§3.1 and throughout] The model name is inconsistent: 'Qwen-3.7-Plus' in §3.1 versus 'Qwen 3.7-Plus' elsewhere. Please harmonize.
  2. [Appendix A.1] PassRate is defined as the fraction of tasks with score at least 0.8, while the main text says 'percentage of fully passed tasks.' Please reconcile the two definitions.
  3. [Appendix A.3] The text refers to 'CUA-Harness' in the Terminal-Bench setup; this appears to be a typo for LongHorizon-Harness.
  4. [Fig. 3] The asterisk legend for externally reported metrics is incomplete. Please mark each externally reported value explicitly.
  5. [§2.3] The executor is called 'fresh-context' but receives prior audit reports referenced by the contract. Please clarify that freshness applies to the raw interaction trajectory, not to all information.
  6. [Reproducibility] The paper would benefit from providing the manager/executor/auditor prompt templates and the task-state schema, or a clear pointer to them in the public repository, since these details are central to reproducing the MEA loop.

Circularity Check

0 steps flagged

No circularity: the paper's claims rest on external benchmark evaluations, not on fitted parameters, self-citation chains, or definitional reductions.

full rationale

LongHorizon-Harness makes an empirical claim: a Manage-Execute-Audit loop with external task state improves long-horizon agent completion. The paper contains no fitted model, no parameter estimation, and no derivation that reduces a prediction to an input. The observed improvements on WeaveBench, Terminal-Bench 2.1, and OSWorld 2.0 are measured against external benchmarks using matched backbones in most comparisons. The same-backbone manager/executor/auditor design and the OSWorld evaluator's qwen3.7-plus configuration are validity concerns about independence and evaluation fairness, not circular derivation: the paper does not define its outcome in terms of those components, nor does it cite its own prior work as the load-bearing justification for its central mechanism. The paper itself flags the key limitation in Appendix B.2.2: 'plausible evidence can be mistaken for completed evidence' and 'a misinterpreted contract can still lead to a confidently verified wrong answer.' This is an acknowledged residual risk in the auditor premise, not a circular step. Self-citations appear in the related-work section (e.g., GPG, SkillClaw, Coevolve), but none is used to justify the harness design or the benchmark gains. Therefore, no enumerated circularity pattern is present, and the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central claim rests on unproved modeling assumptions about auditor reliability and fresh-context sufficiency, plus hand-set budgets and inference settings. No ablation or sensitivity analysis is provided for these choices, so they are free parameters of the empirical result.

free parameters (5)
  • Nmax = 25 maximum MEA rounds = 25
    Chosen as round budget; no sensitivity analysis, affects whether a task can be completed.
  • Executor timeout = 1800 s
    Hand-set per-round budget; success rate is likely budget-dependent.
  • Manager and auditor timeout = 300 s
    Hand-set; too short may truncate verification, too long increases cost.
  • Model inference settings for Qwen runs = temperature=1.0, top_p=0.95, top_k=20, max_tokens=65536, effort=max
    Chosen for Qwen and Claude runs; no ablations are reported.
  • WeaveBench judge configuration = Claude Opus 4.7, AJ_THINKING=medium
    The evaluation judge is itself a model; results depend on this choice.
axioms (4)
  • domain assumption A read-only auditor using the same backbone model can reliably determine task completion and integrity from environment observation.
    Invoked throughout Section 2.4; if the auditor has false negatives or false positives, the external task-state ledger is corrupted and later rounds plan from wrong facts.
  • domain assumption Fresh-context execution with only the current contract, task state, and referenced audit reports carries all information needed for the next subtask.
    Section 2.3 discards raw trajectories; if earlier context not captured in state is necessary, discarding it harms performance.
  • domain assumption The benchmarks' evaluators are valid proxies for real long-horizon task success.
    WeaveBench uses an LLM judge (Claude Opus 4.7), and the OSWorld native evaluator is configured with qwen3.7-plus; all headline numbers depend on these judges.
  • domain assumption Root privileges in the WeaveBench task VMs do not change the matched comparison.
    The paper states official results use a regular user while the paper's runs use root; both matched runs use root, but this setup may make tasks easier than real deployment.
invented entities (2)
  • Explicit task-state ledger (requirements, artifacts, facts with status) no independent evidence
    purpose: External cross-round memory that only audited information can update
    A software construct, not a physical entity; its correctness is not independently observable outside the implementation, and the paper provides no direct measurement of ledger accuracy.
  • Audit report as the only cross-round memory no independent evidence
    purpose: Replace raw trajectory with verified environment evidence
    An internal design element; the claim that audit reports are the only memory is asserted but not separately falsified.

pith-pipeline@v1.3.0-daily-deepseek · 31287 in / 14538 out tokens · 162342 ms · 2026-08-04T17:35:59.986777+00:00 · methodology

0 comments
read the original abstract

Large language model (LLM) agents increasingly undertake long-horizon tasks that require sustained reasoning, tool use, and revision across many interdependent steps. However, existing agent harnesses maintain task execution, task state, and completion assessment within a growing context, making the state difficult to track and allowing incorrect self-assessments to propagate into later decisions. We reformulate long-horizon execution as a task-state management problem and propose LongHorizon-Harness, which maintains the task state explicitly outside execution and updates it only with facts independently verified from the environment. Its Manage-Execute-Audit(MEA) loop uses a manager to maintain the task state and determine the next subtask, a fresh-context executor to perform it, and a read-only auditor to verify the resulting environment state before the next round. A lightweight AgentAdapter supports interchangeable model and harness backends without modifying their native agent loops. LongHorizon-Harness improves Qwen~3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench~2.1, and from 2.8% to 8.3% on OSWorld~2.0. It also raises Claude Opus~4.7 from 20.0% to 34.3% on an OSWorld2.0 subset, demonstrating consistent gains across models, harnesses, and interaction domains.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

24 extracted references · 10 linked inside Pith

  1. [1]

    Agent S: An open agentic frameworkthatusescomputerslikeahuman

    Saaket Agashe, Jiuzhou Han, Shuyu Gan, Jiachen Yang, Ang Li, and Xin Eric Wang. Agent S: An open agentic frameworkthatusescomputerslikeahuman. InThe Thirteenth International Conference on Learning Representations, 2025.https://arxiv.org/abs/2410.08164. Anthropic. Claude computer use. https://www.anthropic.com/news/3-5-models-and-computer-use ,

    Pith/arXiv arXiv 2025

  2. [5]

    Nl2repo-bench: Towards long-horizon repository generation evaluation of coding agents.arXiv preprint arXiv:2512.12730,

    Jingzhe Ding, Shengda Long, Changxin Pu, Huan Zhou, Hongwan Gao, Xiang Gao, Chao He, Yue Hou, Fei Hu, Zhaojian Li, et al. Nl2repo-bench: Towards long-horizon repository generation evaluation of coding agents.arXiv preprint arXiv:2512.12730,

    arXiv

  3. [6]

    Google DeepMind

    doi: 10.20944/preprints202607.1328.v1.https://doi.org/10.20944/preprints202607.1328.v1. Google DeepMind. Introducing the Gemini 2.5 Computer Use model.https://blog.google/innovation-and-ai/mo dels-and-research/google-deepmind/gemini-computer-use-model/,

    arXiv

  4. [7]

    Kelly Hong, Anton Troynikov, and Jeff Huber

    Accessed: 2026-06-22. Kelly Hong, Anton Troynikov, and Jeff Huber. Context rot: How increasing input tokens impacts LLM performance. https://research.trychroma.com/context-rot,

    2026

  5. [8]

    Tree search for llm agent reinforcement learning.arXiv preprint arXiv:2509.21240,

    Yuxiang Ji, Ziyu Ma, Yong Wang, Guanhua Chen, Xiangxiang Chu, and Liaoni Wu. Tree search for llm agent reinforcement learning.arXiv preprint arXiv:2509.21240,

    arXiv

  6. [9]

    Measuring AI ability to complete long software tasks, 2026.https://arxiv.or g/abs/2503.14499

    Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, et al. Measuring AI ability to complete long software tasks, 2026.https://arxiv.or g/abs/2503.14499. Renda Li, Hailang Huang, Fei Wei, Feng Xiong, Yong Wang, and Xiangxiang Chu. Adacurl: Adaptive curriculum reinforcement learni...

    Pith/arXiv arXiv 2026

  7. [10]

    WeaveBench: A long-horizon, real-world benchmark for computer-use agents with hybrid interfaces, 2026.https://arxiv.org/abs/2606.09426

    Wanli Li, Bowen Zhou, Yunyao Yu, Zhou Xu, Yifan Yang, Dongsheng Li, and Caihua Shan. WeaveBench: A long-horizon, real-world benchmark for computer-use agents with hybrid interfaces, 2026.https://arxiv.org/abs/2606.09426. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How lan...

    Pith/arXiv arXiv 2026

  8. [11]

    Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, and Xiangxiang Chu

    doi: 10.1162/tacl_a_00638. Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, and Xiangxiang Chu. Skillclaw: Let skills evolve collectively with agentic evolver.arXiv preprint arXiv:2604.08377,

    Pith/arXiv arXiv

  9. [12]

    Merrill, Alexander G

    Mike A. Merrill, Alexander G. Shaw, Nicholas Carlini, Boxuan Li, Harsh Raj, Ivan Bercovich, et al. Terminal-Bench: Benchmarking agents on hard, realistic tasks in command line interfaces, 2026.https://arxiv.org/abs/2601.11868. Nous Research. Hermes agent CLI.https://hermes-agent.ai/tools/hermes-agent-cli,

    Pith/arXiv arXiv 2026

  10. [13]

    Accessed: 2026-05-05. OpenAI. Introducing Codex.https://openai.com/index/introducing-codex/, May 2025a. Accessed: 2026-05-07. OpenAI. Introducing ChatGPT agent: bridging research and action.https://openai.com/index/introducing-cha tgpt-agent/, July 2025b. Accessed: 2026-05-07. OpenAI. Operator.https://openai.com/research/operator, 2025c. Accessed: 2025-05...

    2026

  11. [14]

    16 Qingnan Ren, Shun Zou, Shiting Huang, Ziao Zhang, Kou Shi, Zhen Fang, Yiming Zhao, Yu Zeng, Qisheng Su, Lin Chen, et al

    Accessed: 2026-06-22. 16 Qingnan Ren, Shun Zou, Shiting Huang, Ziao Zhang, Kou Shi, Zhen Fang, Yiming Zhao, Yu Zeng, Qisheng Su, Lin Chen, et al. Saasbench: Exploring the boundaries of coding agents in long-horizon enterprise saas engineering.arXiv preprint arXiv:2605.17526,

    Pith/arXiv arXiv 2026

  12. [15]

    ScienceBoard: Evaluating multimodal autonomous agents in realistic scientific workflows, 2025.https://arxiv.org/abs/2505.19897

    Qiushi Sun, Zhoumianze Liu, Chang Ma, Zichen Ding, Fangzhi Xu, Zhangyue Yin, Haiteng Zhao, Zhenyu Wu, Kanzhi Cheng, Zhaoyang Liu, et al. ScienceBoard: Evaluating multimodal autonomous agents in realistic scientific workflows, 2025.https://arxiv.org/abs/2505.19897. Yiyou Sun, Xinyang Han, Weichen Zhang, Yuanbo Pang, Tianyu Wang, Yuhan Cao, Yixiao Huang, et...

    Pith/arXiv arXiv 2025

  13. [17]

    Code2world: A gui world model via renderable code generation.arXiv preprint arXiv:2602.09856,

    Yuhao Zheng, Li’an Zhong, Yi Wang, Rui Dai, Kaikui Liu, Xiangxiang Chu, Linyuan Lv, Philip Torr, and Kevin Qinghong Lin. Code2world: A gui world model via renderable code generation.arXiv preprint arXiv:2602.09856,

    arXiv

  14. [18]

    Colorbrowseragent: An intelligent gui agent for complex long-horizon web automation.arXiv preprint arXiv:2601.07262,

    Jiamu Zhou, Jihong Wang, Weiming Zhang, Weiwen Liu, Zhuosheng Zhang, Xingyu Lou, Weinan Zhang, Huarong Deng, and Jun Wang. Colorbrowseragent: An intelligent gui agent for complex long-horizon web automation.arXiv preprint arXiv:2601.07262,

    Pith/arXiv arXiv

  15. [19]

    Ashley, Wenyi Wang, Dmitrii Khizbullin, Yunyang Xiong, Zechun Liu, Ernie Chang, Raghuraman Krishnamoorthi, Yuandong Tian, Yangyang Shi, Vikas Chandra, and Jürgen Schmidhuber

    Mingchen Zhuge, Changsheng Zhao, Dylan R. Ashley, Wenyi Wang, Dmitrii Khizbullin, Yunyang Xiong, Zechun Liu, Ernie Chang, Raghuraman Krishnamoorthi, Yuandong Tian, Yangyang Shi, Vikas Chandra, and Jürgen Schmidhuber. Agent-as-a-Judge: Evaluate agents with agents, 2024.https://arxiv.org/abs/2410.10934. 17 A Detailed Experimental Setup A.1 WeaveBench WeaveB...

    Pith/arXiv arXiv 2024

  16. [20]

    We use the official OSWorld-v2 osworld-v2-2026.06.24 release, including the corresponding task definitions, task assets, Docker VM image, and mocked websites

    consists of 108 professional desktop workflow tasks, with a median human completion time of approximately 1.6 hours. We use the official OSWorld-v2 osworld-v2-2026.06.24 release, including the corresponding task definitions, task assets, Docker VM image, and mocked websites. Experiments are conducted using the standard OSWorld Docker-based VM infrastructu...

    2026

  17. [22]

    Zero” and “Full

    is designed to evaluate agents on challenging and realistic software- engineering and command-line tasks. We use Harbor as the evaluation framework, with the Docker backend providing an isolated execution environment for each task. During evaluation, we preserve the CPU, memory, and environment constraints originally defined for each task. To ensure consi...

    2093

  18. [23]

    ✓✓ Figure 15Data-analysis case

    viewairflowtaskinstancedetails.png: Shows the Task Instance Details panel for extractorders @ slaevenhour with: - Status:… The verifier reviews each screenshot and its metadata, confirming the evidence chain is complete and compliant. ✓✓ Figure 15Data-analysis case. The harness rejects mislabeled evidence, captures the correct Airflow views, and audits wh...

    1920

  19. [150]

    YELLOW:",$0}' | head -30 $ which objdump nm readelf gdb ltrace strace 2>&1; ls /usr/bin/ | grep -iE

    &| (r<100 && g>150)) print}' | head -40 $ awk 'NR>3 {print NR-3, $0}' /app/image.ppm | awk '{r=$2; g=$3; b=$4; if (!( … 80 && b<100) print "YELLOW:",$0}' | head -30 $ which objdump nm readelf gdb ltrace strace 2>&1; ls /usr/bin/ | grep -iE "ghidra|ida|radare|rizin|r2" 2>&1 $ nm /app/mystery 2>&1 | grep -iE " T |sphere|ray|camera|scene|render|write_i … r|h...

    2000

  20. [1080]

    Unlike the main official baselines, which typically rely only on GUI actions, LongHorizon-Harness uses a hybrid tool pool on OSWorld-v2

    Final scoring is performed by the native OSWorldenv.evaluate()procedure. Unlike the main official baselines, which typically rely only on GUI actions, LongHorizon-Harness uses a hybrid tool pool on OSWorld-v2. GUI actions are executed by a computer MCP server inside the VM, supporting desktop operations such as clicking, typing, dragging, and hotkeys. The...

    2000

  21. [2023]

    OSWorld 2.0: Benchmarking computer use agents on long-horizon real-world tasks, 2026.https://arxiv.org/abs/2606.29537

    Mengqi Yuan, Zilong Zhou, Xinzhuang Xiong, Weiming Wu, Jiayang Sun, Jiamin Song, Kaiqian Cui, Tianbao Xie, Bowen Wang, Haoyuan Wu, Yitong Li, Dunjie Lu, Haikong Lu, Qi Zhen, Xinyuan Wang, Tao Yu, et al. OSWorld 2.0: Benchmarking computer use agents on long-horizon real-world tasks, 2026.https://arxiv.org/abs/2606.29537. Chi Zhang, Zhao Yang, Jiaxuan Liu, ...

    Pith/arXiv arXiv 2026

  22. [2024]

    Anthropic

    Accessed: 2025-05-03. Anthropic. Claude Code.https://www.anthropic.com/product/claude-code,

    2025

  23. [2025]

    Anthropic

    Accessed: 2026-06-22. Anthropic. Claude Cowork.https://www.anthropic.com/product/claude-cowork,

    2026

  24. [2026]

    Xiangxiang Chu, Hailang Huang, Xiao Zhang, Fei Wei, and Yong Wang

    Accessed: 2026-06-24. Xiangxiang Chu, Hailang Huang, Xiao Zhang, Fei Wei, and Yong Wang. Gpg: A simple and strong reinforcement learning baseline for model reasoning. InICLR,

    2026