pith. sign in

arxiv: 2606.13643 · v1 · pith:32HLUC33new · submitted 2026-06-11 · 💻 cs.CL

Recursive Agent Harnesses

Elias Lumer , Sahil Sen , Kevin Paul , Vamse Kumar Subbiah This is my paper

Pith reviewed 2026-06-27 06:39 UTC · model grok-4.3

classification 💻 cs.CL
keywords recursive agent harnessharness recursionlong-context reasoningcoding agentsrecursive language modelsagent workflowssubagent spawning
0
0 comments X

The pith

Recursive agent harnesses raise a fixed GPT-5 coding baseline from 71.75 percent to 81.36 percent on long-context tasks.

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

The paper defines Recursive Agent Harness as the use of recursion over full agent units that include filesystem tools, code execution, and planning, rather than bare model calls. It positions this harness recursion as the code-first extension of the model recursion introduced in recursive language models. A parent agent writes and executes scripts that launch subagent harnesses in parallel for fine-grained work and uses structured calls for small subtasks. With the model backbone held fixed at GPT-5, the design lifts performance on the Oolong-Synthetic benchmark from the Codex coding-agent level of 71.75 percent to 81.36 percent across 199 samples and context lengths up to 4 million tokens. The same structure reaches 89.77 percent when paired with Claude Sonnet 4.5.

Core claim

We name and study the pattern between recursive language models and production coding agents that spawn subagents, where the recursive unit is a full agent harness with tools. We call this the Recursive Agent Harness and frame it as harness recursion. A parent agent generates and runs an executable script that spawns subagent harnesses in parallel for fine-grained workloads and uses structured function calls for small subtasks. With the backbone held fixed at GPT-5 to match published baselines, RAH improves the Codex coding-agent baseline from 71.75 percent to 81.36 percent on Oolong-Synthetic, a gain attributable to the harness rather than the model.

What carries the argument

Recursive Agent Harness (RAH), the executable recursive unit consisting of a full agent with filesystem tools, code execution, and planning that a parent agent spawns via generated scripts.

If this is right

  • The harness structure itself, not model scale, accounts for the measured improvement on long-context coding tasks.
  • Parallel spawning of sub-harnesses enables fine-grained division of long-context workloads.
  • Structured function calls inside the harness handle small subtasks without full agent overhead.
  • The same harness design scales to stronger backbones, reaching 89.77 percent with Claude Sonnet 4.5.

Where Pith is reading between the lines

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

  • The approach could be tested on non-coding long-context benchmarks to check whether harness recursion generalizes beyond code generation.
  • Different recursion depths or hybrid model-plus-harness recursion patterns remain open for direct comparison under the same fixed-backbone protocol.

Load-bearing premise

The controlled evaluation with fixed backbone and chosen baseline fully isolates the contribution of the recursive harness structure from differences in prompting, tool access, or other implementation details.

What would settle it

A controlled re-run of the Codex baseline that exactly matches the RAH prompting, tool set, and script-generation format, after which the performance gap disappears.

Figures

Figures reproduced from arXiv: 2606.13643 by Elias Lumer, Kevin Paul, Sahil Sen, Vamse Kumar Subbiah.

Figure 1
Figure 1. Figure 1: The Recursive Agent Harness (RAH). A parent agent selects between code-execution spawning (writing an executable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Oolong Score on Oolong-Synthetic (199 samples), [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
read the original abstract

Recursive language models (RLMs) showed that recursion over model calls is an effective strategy for long-context reasoning, and production coding agents have begun to write code that spawns subagents at scale, most recently in Anthropic's dynamic workflows. We name and study the pattern between these two lines of work, where the recursive unit is a full agent harness with filesystem tools, code execution, and planning rather than a model call with no tools. We call this the Recursive Agent Harness (RAH) and frame it as harness recursion, the code-first extension to the model recursion of RLMs. A parent agent generates and runs an executable script that spawns subagent harnesses in parallel for fine-grained workloads and uses structured function calls for small subtasks. We provide a controlled evaluation on long-context reasoning. With the backbone held fixed at GPT-5 to match the published Codex and RLM baselines, RAH improves the Codex coding-agent baseline from 71.75% to 81.36% on Oolong-Synthetic (199 samples, 13 context-length buckets up to 4M tokens), a gain attributable to the harness rather than the model. With a stronger backbone, Claude Sonnet 4.5, the same design reaches 89.77%.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Recursive Agent Harnesses (RAH) as a code-first extension of recursive language models, in which a parent agent generates executable scripts that spawn parallel subagent harnesses equipped with filesystem tools, code execution, and structured function calls. With the model backbone fixed at GPT-5, the paper reports that RAH raises performance on the Oolong-Synthetic benchmark (199 samples, 13 context-length buckets up to 4 M tokens) from the Codex coding-agent baseline of 71.75 % to 81.36 %, attributing the 9.61 pp gain to the harness recursion rather than the underlying model; a stronger backbone (Claude Sonnet 4.5) reaches 89.77 %.

Significance. If the controlled evaluation is reproducible, the work usefully bridges recursive language models and production-scale coding agents by demonstrating that recursion at the level of full agent harnesses can improve long-context reasoning performance. The explicit use of a fixed backbone to match published baselines is a methodological strength that supports attribution claims.

major comments (2)
  1. [Abstract / Evaluation] Abstract and evaluation section: the claim that the 9.61 pp gain on Oolong-Synthetic is 'attributable to the harness rather than the model' is load-bearing for the central contribution, yet the manuscript supplies no details on variance across runs, exact reproduction of the Codex baseline code, data-exclusion rules, or statistical tests. Without these, the isolation of the recursive-harness effect cannot be verified.
  2. [Evaluation methodology] Evaluation methodology: the controlled comparison with fixed GPT-5 backbone assumes that all non-recursive implementation details (system prompts, tool schemas, filesystem access, parallel-execution handling, and context management) are identical between RAH and the Codex baseline. The text does not state that such equivalence was explicitly verified or that the baseline was re-run inside the authors' harness framework.
minor comments (2)
  1. [Abstract] The abstract mentions '13 context-length buckets' but does not indicate how bucket boundaries or sample distribution are defined; a brief table or sentence would improve clarity.
  2. [Introduction] The paper cites prior RLM and Anthropic workflow work but could add a short related-work paragraph distinguishing harness recursion from model-level recursion more explicitly.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the need for greater methodological transparency in our controlled evaluation. We address each major comment below and commit to revisions that strengthen the verifiability of the attribution claim without altering the core experimental design.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and evaluation section: the claim that the 9.61 pp gain on Oolong-Synthetic is 'attributable to the harness rather than the model' is load-bearing for the central contribution, yet the manuscript supplies no details on variance across runs, exact reproduction of the Codex baseline code, data-exclusion rules, or statistical tests. Without these, the isolation of the recursive-harness effect cannot be verified.

    Authors: We agree that the manuscript would benefit from explicit documentation of these elements to support reproducibility. The reported 71.75% Codex baseline is taken directly from the published results to enforce the fixed GPT-5 backbone control, consistent with standard practice when matching prior work. In revision we will: (1) add a dedicated reproducibility subsection describing the baseline matching protocol, (2) specify any data-exclusion rules applied to the 199-sample Oolong-Synthetic set, (3) report that results reflect single runs aligned with the original baseline publications, and (4) include variance or statistical test results if additional runs were conducted during development (or explicitly note their absence as a limitation). revision: yes

  2. Referee: [Evaluation methodology] Evaluation methodology: the controlled comparison with fixed GPT-5 backbone assumes that all non-recursive implementation details (system prompts, tool schemas, filesystem access, parallel-execution handling, and context management) are identical between RAH and the Codex baseline. The text does not state that such equivalence was explicitly verified or that the baseline was re-run inside the authors' harness framework.

    Authors: The evaluation deliberately uses the published Codex numbers rather than re-implementing the baseline inside the RAH framework, because the central control variable is the model backbone itself. This isolates the contribution of harness recursion (executable script spawning of subagents) while holding the underlying model fixed. We will revise the evaluation methodology section to state this design choice explicitly, describe how non-recursive components were aligned with published descriptions to the extent possible, and clarify that full re-implementation of the baseline within our harness was not performed as it would confound the fixed-backbone comparison. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical result with fixed backbone

full rationale

The paper's central claim is a measured empirical improvement (71.75% to 81.36%) on Oolong-Synthetic with GPT-5 backbone held fixed to match baselines. No derivation, equations, fitted parameters, or self-citations are invoked to produce this number; the gain is presented as the direct output of the controlled experiment. No load-bearing step reduces by construction to the inputs, satisfying the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper introduces an architectural pattern resting on one domain assumption about the utility of code-mediated recursion for long-context tasks and one newly defined entity; no free parameters or mathematical axioms are stated.

axioms (1)
  • domain assumption Long-context reasoning tasks benefit from parallel decomposition into subagent harnesses spawned via executable code.
    Implicit premise underlying the choice of RAH design and the controlled evaluation.
invented entities (1)
  • Recursive Agent Harness (RAH) no independent evidence
    purpose: A full agent system with filesystem tools, code execution, and planning that can be recursively spawned by a parent agent via code.
    Newly named and studied pattern extending model recursion.

pith-pipeline@v0.9.1-grok · 5753 in / 1203 out tokens · 25269 ms · 2026-06-27T06:39:52.224336+00:00 · methodology

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

34 extracted references · 18 linked inside Pith

  1. [1]

    Anonymous. 2026. AGENTHIVE: A Composable Multi-Agent Framework with First-Class Delegation. ACL ARR 2026 January submission. openre- view.net/forum?id=BYiwNYYixO. Authors are anonymous (double-blind). Re- place with deanonymized authors upon publication

    2026

  2. [2]

    Anthropic. 2026. Orchestrate Subagents at Scale with Dynamic Workflows. https: //code.claude.com/docs/en/workflows. Claude Code documentation, research preview. Accessed 2026-06-06

    2026

  3. [3]

    Amanda Bertsch, Adithya Pratapa, Teruko Mitamura, Graham Neubig, and Matthew R. Gormley. 2025. Oolong: Evaluating Long Context Reasoning and Ag- gregation Capabilities.arXiv preprint arXiv:2511.02817(2025). arXiv:2511.02817

    arXiv 2025

  4. [4]

    Weili Cao, Xunjian Yin, Bhuwan Dhingra, and Shuyan Zhou. 2026. Coding Agents are Effective Long-Context Processors.arXiv preprint arXiv:2603.20432 (2026). arXiv:2603.20432

    arXiv 2026

  5. [5]

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024. AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors. InInternational Conference on Learning Representations. arXiv:2308.10848

    Pith/arXiv arXiv 2024

  6. [6]

    Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

    Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2024. SWE-bench: Can Language Models Resolve Real-World GitHub Issues?. InInternational Conference on Learning Representa- tions. arXiv:2310.06770

    Pith/arXiv arXiv 2024

  7. [7]

    Mahoney, Kurt Keutzer, and Amir Gholami

    Sehoon Kim, Suhong Moon, Ryan Tabrizi, Nicholas Lee, Michael W. Mahoney, Kurt Keutzer, and Amir Gholami. 2024. An LLM Compiler for Parallel Function Calling. InInternational Conference on Machine Learning. arXiv:2312.04511

    arXiv 2024

  8. [8]

    LangChain. 2022. LangChain. https://www.langchain.com/

    2022

  9. [9]

    Daniel Freeman, Sara de Jong, Arthur Gretton, and Mandar Joshi

    Jinhyuk Lee, Anthony Ing, Zhuyun Dai, Sharan Narang, Kelvin Guu, Benjamin Pitchford, Maxwell Chang, Steven Wan, Ryutaro Tanno, C. Daniel Freeman, Sara de Jong, Arthur Gretton, and Mandar Joshi. 2024. Can Long-Context Language Models Subsume Retrieval, RAG, SQL, and More?. InProceedings of the 62nd Annual Meeting of the Association for Computational Lingui...

    arXiv 2024

  10. [10]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the Middle: How Language Models Use Long Contexts.Transactions of the Association for Computational Linguistics 12 (2024), 157–173. arXiv:2307.03172

    Pith/arXiv arXiv 2024

  11. [11]

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianhao Fu, Yuxiao Liu, Zihan Yao, Rui Zhang, Jie Jia, Jie Tang, Yuxiao Liu, and Yuxiao Dong. 2024. AgentBench: Evaluating LLMs as Agents. InInternational Conferenc...

    Pith/arXiv arXiv 2024

  12. [12]

    Elias Lumer, Anmol Gulati, Faheem Nizar, Dzmitry Hedroits, Atharva Mehta, Henry Hwangbo, Vamse Kumar Subbiah, Pradeep Honaganahalli Basavaraju, and James A Burke. 2026. Tool and agent selection for large language model agents in production: A survey. In2026 IEEE Conference on Artificial Intelligence (CAI). IEEE, 701–708

    2026

  13. [13]

    Elias Lumer, Faheem Nizar, Akshaya Jangiti, Kevin Frank, Anmol Gulati, Man- dar Phadate, and Vamse Kumar Subbiah. 2026. Don’t Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks.arXiv preprint arXiv:2601.06007(2026)

    arXiv 2026

  14. [14]

    Grégoire Mialon, Roberto Dessì, Maria Lomeli, Christoforos Nalmpantis, Ram Pasunuru, Roberta Raileanu, Baptiste Rozière, Timo Schick, Jane Dwivedi-Yu, Asli Celikyilmaz, Edouard Grave, Yann LeCun, and Thomas Scialom. 2023. Augmented Language Models: a Survey.Transactions on Machine Learning Research(2023). arXiv:2302.07842

    Pith/arXiv arXiv 2023

  15. [15]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. 2021. WebGPT: Browser-Assisted Question-Answering with Human Feedback.arXiv preprint...

    Pith/arXiv arXiv 2021

  16. [16]

    Avanika Narayan, Dan Biderman, Sabri Eyuboglu, Avner May, Scott Linderman, James Zou, and Christopher Re. 2025. Minions: Cost-efficient Collaboration Between On-device and Cloud Language Models. arXiv:2502.15964 [cs.LG] https://arxiv.org/abs/2502.15964

    arXiv 2025

  17. [17]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. 2023. MemGPT: Towards LLMs as Operating Systems.arXiv preprint arXiv:2310.08560(2023)

    Pith/arXiv arXiv 2023

  18. [18]

    Yujia Qin, Shihao Liang, Yining Ye, Kunliang Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun

  19. [19]

    InInternational Conference on Learning Representations

    ToolLLM: Facilitating Large Language Models to Master 16000+ Real-World APIs. InInternational Conference on Learning Representations. arXiv:2307.16789

    Pith/arXiv arXiv

  20. [20]

    Amartya Roy, Rasul Tutunov, Xiaotong Ji, Matthieu Zimmer, and Haitham Bou- Ammar. 2026. The Y-Combinator for LLMs: Solving Long-Context Rot with 𝜆-Calculus.arXiv preprint arXiv:2603.20105(2026). Submitted 20 Mar 2026

    arXiv 2026

  21. [21]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Systems. arXiv:2302.04761

    Pith/arXiv arXiv 2023

  22. [22]

    Sahil Sen, Akhil Kasturi, Elias Lumer, Anmol Gulati, and Vamse Kumar Subbiah

  23. [23]

    arXiv preprint arXiv:2605.15184(2026)

    Is Grep All You Need? How Agent Harnesses Reshape Agentic Search. arXiv preprint arXiv:2605.15184(2026). https://arxiv.org/abs/2605.15184

    Pith/arXiv arXiv 2026

  24. [24]

    Sahil Sen, Elias Lumer, Anmol Gulati, and Vamse Kumar Subbiah. 2026. Chronos: Temporal-aware conversational agents with structured event retrieval for long- term memory.arXiv preprint arXiv:2603.16862(2026)

    arXiv 2026

  25. [25]

    Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L

    Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths

  26. [26]

    arXiv:2309.02427

    Cognitive Architectures for Language Agents.Transactions on Machine Learning Research(2024). arXiv:2309.02427

    Pith/arXiv arXiv 2024

  27. [27]

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An Open-Ended Embodied Agent with Large Language Models.arXiv preprint arXiv:2305.16291(2023)

    Pith/arXiv arXiv 2023

  28. [28]

    Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. 2024. Executable Code Actions Elicit Better LLM Agents. arXiv:2402.01030 [cs.CL] https://arxiv.org/abs/2402.01030

    arXiv 2024

  29. [29]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. InAdvances in Neural Information Processing Systems. arXiv:2201.11903

    Pith/arXiv arXiv 2022

  30. [30]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. 2023. AutoGen: Enabling Next- Gen LLM Applications via Multi-Agent Conversation. arXiv:2308.08155 [cs.AI] https://arxiv.org/abs/2308.08155

    Pith/arXiv arXiv 2023

  31. [31]

    Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press

    John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. 2024. SWE-agent: Agent-Computer Inter- faces Enable Automated Software Engineering. InAdvances in Neural Information Processing Systems. arXiv:2405.15793

    Pith/arXiv arXiv 2024

  32. [32]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. InAdvances in Neural Information Processing Systems. arXiv:2305.10601

    Pith/arXiv arXiv 2023

  33. [33]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Mod- els. InInternational Conference on Learning Representations. arXiv:2210.03629

    Pith/arXiv arXiv 2023

  34. [34]

    subtask", prompt=f

    Alex L. Zhang, Tim Kraska, and Omar Khattab. 2026. Recursive Language Models. arXiv:2512.24601 [cs.AI] https://arxiv.org/abs/2512.24601 6 Lumer et al. A Parent Agent Prompt The parent agent runs a general-purpose harness prompt. It con- tains no task- or benchmark-specific instructions, and it does not tell the agent what to extract, how to score, or how ...

    Pith/arXiv arXiv 2026