acceptodds
Under review as a conference paper at ICLR 2027

EvoHarness-RL: Learning Runtime Harness Coordination for Self-Evolving Agents

Abstract

Long-horizon LLM agents increasingly rely on external execution support to maintain state, track progress, recover from failures, and reuse experience across extended interactions. Yet existing harnesses and their use are often tailored to environments and controlled through prompts, heuristics, or system-specific rules, making agent and harness coordination difficult to jointly optimize. We introduce EVOHARNESS-RL, a unified framework that separates environment-specific harness implementations from a shared policy-facing interface. EVOHARNESS-RL organizes external support into a Belief, Progress, and Experience (BPE) workspace and exposes four compact harness actions for accessing and updating this state. We first instantiate BPE as an inference-time scaffold and then make harness coordination learnable through supervised initialization followed by cost-aware GRPO. Across heterogeneous long-horizon tasks, EVOHARNESS-BASE improves the average success rate of frontier models by 10.0 percentage points, while EVOHARNESS-RL outperforms the strongest open-source baseline by 8.5 percentage points, together with higher RL rollout efficiency and stronger generalization to unseen tasks. Our analyses show that training gradually shifts agents from frequent scaffold use toward selective, environment-dependent harness access as the policy becomes more capable, while the external workspace continues to evolve and refine itself to better support task execution and generalization. Together, these results show that long-horizon agents benefit not only from external scaffolding itself, but also from learning how and when to coordinate with external support as part of a cost-aware policy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Agent Harness

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ujQmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This article presents EvoHarness-RL, which unifies the agent's memory, state tracking, and progress management into a BPE framework, and trains the agent to autonomously use these auxiliary functions through SFT and GRPO. Experiments demonstrate improvements in task success rates and training efficiency across three benchmarks.

Critical strengths

The research question is reasonable and meaningful, and the method is simple and makes sense..

The experiment is complete and convincing, demonstrating the effectiveness of cross-task generalization.

Critical weaknesses

The novelty of the method is relatively limited. There has been a lot of existing work on external memory, in-context tool using, and RL training.

The authors claim to provide a unified Harness framework, but what is unified is mainly the interface, rather than the actual implementation. This may limit the generality of the framework.

Questions for authors

See Weaknesses

3 comments
AuthorsOpenReview ↗

Official Reply to Reviewer ujQm (Part 1)

We sincerely thank the reviewer for recognizing the significance of our research question and the simplicity and soundness of our method. We also appreciate your thoughtful feedback. We have made every effort to address your concerns through point-by-point clarifications. Our responses to the weaknesses are provided below.


W1: Novelty Compared with Existing Work

W1: The novelty of the method is relatively limited. There has been a lot of existing work on external memory, in-context tool use, and RL training.

Response:

We agree that external memory, tool use, and agent RL have received substantial attention in recent research. Much of the existing work emphasizes either harness-side mechanisms, such as designing, retrieving, and evolving external knowledge and reusable skills (e.g., Reflexion [1], ExpeL [2], ACE [3], SkillOS [4]), or policy-side optimization, including RL training with external skills (e.g., SkillRL [5]). However, our work focuses on the coordination between agent policies and their runtime harnesses, a relatively underexplored research direction at their intersection, to systematically investigate their complementary roles and co-evolution dynamics across heterogeneous environments.

To address this, we introduce a unified BPE abstraction and a shared coordination protocol, making harness use an explicit, learnable policy decision. This enables us to systematically investigate and optimize the evolving division of labor between what agents learn internally and what they delegate to external harness support.

Our experiments demonstrate both practical improvements and new insights into policy–harness co-evolution across heterogeneous environments. We observe harness annealing, where policies progressively reduce their reliance on routine harness operations during training, alongside environment-dependent coordination strategies reflecting different demands for external support. Meanwhile, the harness facilitates policy learning by improving RL rollout efficiency, while accumulated experience progressively enhances its support capabilities. Together, these findings reveal how agent policies and external harnesses can complement and adapt to each other throughout learning.

References

[1] Reflexion: Language Agents with Verbal Reinforcement Learning

[2] ExpeL: LLM Agents Are Experiential Learners

[3] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

[4] SkillOS: Learning Skill Curation for Self-Evolving Agents

[5] SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning

AuthorsOpenReview ↗

Official Reply to Reviewer ujQm (Part 2)

W2: Scope and Generality of Harness Unification

W2: The authors claim to provide a unified Harness framework, but what is unified is mainly the interface, rather than the actual implementation. This may limit the generality of the framework.

Response:

Our objective is not to build a fully plug-and-play universal harness system, but to unify the policy-facing harness coordination interface across heterogeneous environments.

Specifically, BPE abstracts heterogeneous underlying harness support into three shared functional roles (Belief, Progress, and Experience), standardizes the interaction protocol between agents and harnesses, and provides a common learning framework for optimizing their coordination. This unification turns harness coordination into a shared, learnable decision problem, enabling us to systematically investigate the complementary roles and co-evolution dynamics of policies and harnesses across heterogeneous environments, including general harness annealing patterns, environment-dependent coordination strategies, and the cross-domain transferability of learned coordination strategies.

We intentionally preserve environment-specific low-level implementations to maintain flexibility and extensibility. Environment adapters allow existing domain-specific and off-the-shelf support tools to be integrated into the shared BPE interface without modifying the coordination protocol. Meanwhile, core implementations, including BPE workspace management, Experience evolution, and the training pipeline, are reused across environments. As quantified in our response to Reviewer X7qy, only 24.1–34.3% of the core harness code needs to be implemented for each environment. This separation enables BPE to accommodate heterogeneous agentic tasks while minimizing duplicated implementation effort.

To further evaluate generality, we additionally instantiate BPE on SWE-bench Verified in our response to Reviewer jhgw, extending our original three benchmarks to challenging coding tasks. EvoHarness-Base improves a closed-source frontier model (Claude) by 4.4 percentage points without fine-tuning, while EvoHarness-RL achieves a 6.2-point improvement over the reported Qwen3-8B baseline, exceeding the GRPO result by 5.0 points. These gains further support BPE's applicability across heterogeneous agentic tasks.


Happy to have further discussion!

Thank you again for recognizing the strengths of our work and for your constructive feedback. We hope our responses address your concerns, and we are happy to discuss if you have any further questions!

AuthorsOpenReview ↗

Looking forward to your reply : )

Dear Reviewer ujQm,

Thank you for your thoughtful review and positive assessment of our work! We truly appreciate your constructive feedback.

With approximately one day remaining in the discussion period, we wanted to gently follow up and ask whether our responses have adequately addressed your concerns. We have carefully clarified the scope of our work, provided quantitative evidence, and presented new results on an additional agentic task. If any part would benefit from further clarification, we would be glad to elaborate.

We sincerely appreciate the time and effort you have devoted to evaluating our paper, and we look forward to your reply!

With our best wishes,
#26876 Authors

Reviewer_jhgwOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces EVOHARNESS-RL, a unified framework designed to optimize the runtime coordination between long-horizon LLM agents and their external execution supports (referred to as "harnesses," such as memory, state trackers, and tools). Existing harnesses are typically environment-specific and manually controlled via prompts or heuristics, making them difficult to jointly optimize with the agent's policy. To solve this, EVOHARNESS-RL abstracts heterogeneous harnesses into a shared BPE (Belief, Progress, and Experience) workspace and exposes four compact meta-actions (track, commit, recall, note). By employing supervised initialization followed by cost-aware reinforcement learning (GRPO) with a diversity-annealing schedule, the framework enables the agent to learn when and how to selectively rely on external scaffolding, gradually internalizing routine decisions as its internal capabilities improve.

Critical strengths

The paper's core strength lies in its unified BPE abstraction, which recasts heterogeneous external scaffolding as a learnable coordination problem, enabling cost-aware reinforcement learning to guide agents toward selective, context-dependent harness invocation—a phenomenon the authors term "harness annealing"—and yielding substantial gains in task success, sample efficiency, and cross-domain generalization across long-horizon benchmarks.

Critical weaknesses

The evaluation is restricted to three relatively constrained benchmarks (ALFWorld, WebShop, WebArena) where information retrieval and state tracking dominate. The paper lacks validation on tasks requiring complex code generation, long-horizon mathematical reasoning, or multi-agent collaboration, where the BPE abstraction might fall short.

Questions for authors

See Weaknesses

2 comments
AuthorsOpenReview ↗

Official Reply to Reviewer jhgw

We sincerely thank the reviewer for the positive evaluation of our work and the very valuable suggestions. We have made every effort to address your concerns through additional experiments and clarifications. Our response to the weakness is provided below.


W1: Generalization Beyond State Tracking and Information Retrieval Benchmarks

W1: The evaluation is restricted to three relatively constrained benchmarks (ALFWorld, WebShop, WebArena) where information retrieval and state tracking dominate. The paper lacks validation on tasks requiring complex code generation, long-horizon mathematical reasoning, or multi-agent collaboration, where the BPE abstraction might fall short.

Response:

Our framework primarily targets long-horizon, stateful agentic tasks involving iterative planning, execution, and feedback, where external state, progress tracking, and reusable experience complement the agent's internal reasoning.

To address this concern, we made additional efforts to instantiate and evaluate EvoHarness-Base and EvoHarness-RL on a challenging coding benchmark, SWE-bench Verified, which requires code understanding, debugging, modification, and testing. We instantiate BPE for coding workflows as follows:

  • Belief (B): Maintains repository states, debugging hypotheses, and their verification status, grounded in code inspection, execution outputs, and test results.
  • Progress (P): Tracks completed and pending subgoals across investigation, implementation, and testing, including their dependencies.
  • Experience (E): Stores and retrieves reusable cross-episode knowledge, including debugging, editing, and testing strategies.

Additional Results on SWE-bench Verified:

Model / MethodResolution RateImprovement
ReAct (Claude)67.0%—
+ EvoHarness-Base71.4%+4.4 pp
Qwen3-8B5.0%—
+ GRPO6.2%+1.2 pp
+ EvoHarness-RL11.2%+6.2 pp

Overall, EvoHarness-Base improves the resolution rate of a frontier model by 4.4 percentage points without policy fine-tuning, while EvoHarness-RL achieves 11.2% resolution rate with Qwen3-8B, outperforming GRPO by 5.0 points.

These results provide additional evidence for the flexibility and applicability of the BPE abstraction across heterogeneous agentic tasks. While extending BPE to long-horizon mathematical reasoning and multi-agent collaboration remains an interesting direction for future work, we see several promising possibilities. For mathematical reasoning, external verifiers and symbolic tools could ground Belief in verified intermediate reasoning states. For multi-agent collaboration, we are particularly interested in exploring shared and local BPE workspaces, where agents maintain their own local Belief, Progress, and Experience while exchanging information through shared states and knowledge. Learning when and how to interact with these workspaces could enable more efficient and effective multi-agent systems.


Happy to have further discussion!

Thank you again for recognizing the strengths of our work and for your constructive feedback. We hope our responses address your concerns, and we are happy to discuss if you have any further questions!

AuthorsOpenReview ↗

Looking forward to your reply : )

Dear Reviewer jhgw,

Thank you again for your very thoughtful and supportive review. We truly appreciate your positive assessment of our work!

As we approach the end of the discussion period, we wanted to gently follow up and ask whether our response and additional results have adequately addressed your concerns. If any part would benefit from further clarification, we would be happy to elaborate.

We truly value your thoughtful feedback and are grateful for the time and care you have devoted to evaluating our work. Looking forward to hearing from you :)

Best regards,
#26876 Authors

Reviewer_X7qyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents EvoHarness-RL, a framework for training an agent to take meta-actions for harness. It selectively interact with an external environment abstraction layer, including Belief, Progress, and Experience workspace. The model operates in a joint action space containing normal tool calls and four BPE meta-actions: track, commit, recall, and note. The method uses teacher traces to initialize this enlarged action space and subsequently applies cost-aware GRPO. The paper evaluates both an always-on BPE scaffold, EvoHarness-Base, and learned BPE coordination, EvoHarness-RL, on ALFWorld, WebShop, and WebArena. Based on evaluation results, improvements are clear. The paper also show component ablations, reward ablations, unseen-task evaluation, cross-backbone experiments, cross-environment initialization, and analyses of harness-use dynamics.

Critical strengths
  • A unified framework to handle diverse environments' harness is an interesting and well-motivated problem. Treating retrieval and external-state updates as model actions is an interesting alternative.

  • Evaluation shows strong performance. The method reports substantial improvements on ALFWorld, WebShop and WebArena.

Critical weaknesses
  • The central “unified harness” claim is not supported. The intro defines a harness broadly to include tools, verifiers, execution, states, click control, test runners, and code search. However, the formal harness action space contains only track, commit, recall, and note, which access or update Belief, Progress, and Experience. Ordinary execution, such as clicking or navigation, remains in a separate environment action space. The proposed method only provides a common interface for limited meta-actions, but does not unify the heterogeneous execution support used to motivate the paper. This mismatch materially overstates the scope and novelty of the contribution.

  • Environment-specific adapters still preserve harness heterogeneity. Each environment still requires manually designed state representations, extraction logic, update behavior, and operation semantics.

  • The efficiency test and evaluation protocols need details. The reported rollout efficiency counts environment and BPE actions but excludes teacher generation, experience consolidation, retrieval, adapter processing, added context tokens, and wall-clock computation. Also, the paper should clearly state whether the persistent experience store is frozen during evaluation and whether one evaluation episode can influence later episodes.

Questions for authors
  1. What is the precise definition of harness used in the paper? Does it include environment tool calls, as suggested by the introduction and Figure 1, or does it refer only to the BPE workspace?

  2. If normal environment actions such as clicking and navigation remain in (A_{\mathrm{env}}), in what sense does BPE unify heterogeneous external execution support?

  3. Can the authors quantify whether BPE reduces the effort required to support a new environment compared with implementing an ordinary environment-specific state tracker and memory system?

6 comments
AuthorsOpenReview ↗

Official Reply to Reviewer X7qy (Part 1)

We sincerely thank the reviewer for providing detailed feedback. To the best of our effort, we carefully address all concerns through point-by-point clarifications. We organize our responses to both the critical weaknesses and questions as follows.


W1 & Q2: Scope of the “Unified Harness” Claim and Relation to Heterogeneous Agent Execution

W1: The proposed method only provides a common interface for limited meta-actions, but does not unify the heterogeneous execution support used to motivate the paper.

Q2: If normal environment actions such as clicking and navigation remain in (A_{\mathrm{env}}), in what sense does BPE unify heterogeneous external execution support?

Response:

Our unification targets the policy-facing harness coordination interface, enabling us to study and optimize how agent policies interact with external support across different environments. In its environment-specific instantiations, this shared high-level interface can integrate domain-specific and off-the-shelf low-level harness components. A potential source of confusion is that Aenv\mathcal{A}_{\mathrm{env}} consists of the agent's task-execution actions defined by each environment (e.g., “click” in WebShop and “go to” in ALFWorld), rather than the underlying harness support functionalities we aim to unify (e.g., webpage-state parsers and object-state and location trackers). Our framework preserves these native environment actions while unifying policy–harness coordination through high-level functional abstractions and a shared action protocol. To clarify this distinction, we separate three layers:

  1. Native task-execution actions (Aenv\mathcal{A}_{\mathrm{env}}; environment-specific, BPE-independent). These are the original action spaces available to the policy agent, as defined by each agentic benchmark, including the “clicking” and “navigation” actions mentioned by the reviewer. They are native task-execution actions rather than external support functionalities provided by the harness. Examples include “go to” / “take” in ALFWorld, “search” / “click” in WebShop, and “click” / “goto” / “type” in WebArena. Their action spaces and execution semantics are inherited directly from the respective benchmarks, as in other agent methods, and remain unchanged regardless of whether BPE is used.

  2. Low-level BPE Harness Instantiations via Environment Adapters (including environment-specific tools). Different environments require different underlying support mechanisms and tools, such as object-state and location trackers in ALFWorld, product-constraint trackers in WebShop, and webpage-state trackers in WebArena. These environment-specific components are organized according to the shared BPE functional roles and operate in the background to provide concrete support for Belief, Progress, and Experience.

  3. Policy-facing harness coordination Interface (BPE workspace, Abpe\mathcal{A}_{\mathrm{bpe}}; shared, unified). This is the layer we unify. BPE abstracts heterogeneous harness support into three functional roles (Belief, Progress, and Experience) and exposes shared meta-actions (“track”, “commit”, “recall”, “note”) for accessing and updating external state. Environment-specific adapters ground these operations in the underlying implementations, while the policy learns to coordinate BPE actions with native environment actions under a shared interaction budget.

Why unify the policy-facing harness coordination interface? Existing harnesses differ in both their underlying implementations and hand-crafted interaction rules, making it difficult to identify generalizable patterns of coordination between agents and harnesses. By standardizing the policy-facing interface while preserving environment-specific implementations, BPE turns harness coordination into a shared, learnable decision problem, enabling systematic study of how the division of labor between policies and harnesses evolves across environments. Our experiments reveal harness annealing as a common phenomenon, alongside environment-specific coordination patterns. Cross-environment transfer further suggests that learned coordination strategies can facilitate faster adaptation to new environments.

Revision: We clarified in Section 2.2 that Aenv\mathcal{A}_{\mathrm{env}} refers to the native task-execution actions defined by each environment.

AuthorsOpenReview ↗

Official Reply to Reviewer X7qy (Part 2)

Q1: Clarifying the Definition and Scope of the Harness

Q1: What is the precise definition of harness used in the paper? Does it include environment tool calls, as suggested by the introduction and Figure 1, or does it refer only to the BPE workspace?

Response: In our response to W1 & Q2, we distinguish three layers to clarify the scope of harness unification. The harness encompasses the high-level BPE workspace (Belief, Progress, and Experience) and its environment-specific low-level instantiations through environment adapters, which incorporate the underlying support tools and modules illustrated in Figure 1. Each environment's domain-specific or off-the-shelf components are organized and mapped to the corresponding BPE functional roles. For example, webpage parsers provide page-state information to support Belief in WebArena, while object-state and location trackers in ALFWorld maintain information about object states and spatial relations. Similarly, memory management modules support Experience by storing and retrieving reusable task experiences. These underlying tools and harness components provide concrete support for maintaining externalized beliefs, tracking task progress, and managing experiences, while environment adapters map their functionalities and outputs to the corresponding BPE workspaces. The agent can then access and coordinate these capabilities through the same shared BPE interface and meta-action protocol across environments, while native environment actions remain separate.

Revision: We revised the Figure 1 caption to clarify how environment adapters ground and map environment-specific support tools and modules into the shared BPE workspace.


W2 & Q3: Environment-Specific Adapters, Implementation Reuse, and Extensibility

W2: Each environment still requires manually designed state representations, extraction logic, update behavior, and operation semantics.

Q3: Can the authors quantify whether BPE reduces the effort required to support a new environment?

Response:

Our objective is not to build a fully plug-and-play universal harness system that directly supports every domain, but rather to unify the policy-facing harness coordination interface to systematically study the complementary roles and co-evolution dynamics of agent policies and harnesses across environments.

  • Why retain environment-specific adapters?

Different environments naturally require different forms of harness support and often already provide mature, off-the-shelf state-tracking mechanisms and grounding tools. Rather than reinventing these components, we use environment adapters to integrate existing domain-specific and off-the-shelf support modules into the shared BPE interface. This design preserves the flexibility and extensibility needed to support heterogeneous agentic environments and tasks, allowing new tools and more sophisticated harness mechanisms to be incorporated without modifying the shared policy-facing coordination protocol.

  • Implementation Reuse and Engineering Effort.

We aim to minimize engineering effort by reusing shared harness infrastructure across environments in our implementations. Our shared core includes BPE workspace management, the common coordination interface, Experience consolidation and evolution mechanisms, and the SFT and cost-aware GRPO training pipeline. Environment-specific adaptations primarily involve Belief representations, state extraction, and integration with underlying support tools, avoiding the need to rebuild the entire harness and optimization infrastructure.

To quantify implementation reuse, we measure the core harness code reuse ratio across three environments, defined as the proportion of shared core source lines of code (SLOC) in each environment's harness implementation.

EnvironmentCode Reuse Ratio
ALFWorld70.2%
WebShop75.9%
WebArena65.7%

Across all three environments, 65.7–75.9% of the core harness implementation is shared, leaving only 24.1–34.3% environment-specific adapter code. This demonstrates substantial implementation reuse and reduces the need to develop a complete harness system from scratch when supporting new environments.

AuthorsOpenReview ↗

Official Reply to Reviewer X7qy (Part 3)

W3: Comprehensive Efficiency Accounting and Evaluation Protocol Clarification

W3: The efficiency test and evaluation protocols need details.

Response:

We thank the reviewer for raising this point. To complement our action-level rollout efficiency analysis, we conducted an additional wall-clock evaluation, showing that EvoHarness-RL reduces training time to reach the same success rate even after accounting for harness-related execution overhead.

Additional Wall-Clock Efficiency Analysis. Extending the rollout efficiency analysis in Appendix B.2, we compare EvoHarness-RL with SFT+GRPO on ALFWorld under the same hardware configuration (8 × NVIDIA H200 GPUs). The table reports the cumulative RL trainer-step wall-clock time to the first validation checkpoint reaching each target success rate, along with estimated GPU rental savings based on an illustrative node price of $32/hour.

Target SRSFT+GRPO UpdatesEvoHarness-RL UpdatesSFT+GRPO Time (h)EvoHarness-RL Time (h)Time ReductionEst. Cost Saving
85%55257.374.9632.7%$77.24
90%70459.378.578.5%$25.49
93%1455019.049.4850.2%$305.85

These wall-clock measurements already account for harness-related runtime overhead, including rollout collection, policy optimization, retrieval and adapter processing, synchronous experience consolidation, blocking API calls and retries, and the processing overhead of additional context tokens along the training critical path. Despite these additional costs, EvoHarness-RL reaches the same performance thresholds with fewer updates and lower cumulative training time.

Regarding teacher generation, both methods use the same number of SFT trajectories. The additional teacher inference cost for EvoHarness-RL is estimated at $16.34 in total, substantially smaller than the $305.85 estimated GPU rental savings at 93% SR.

Evaluation Protocol. Details of experience consolidation during evaluation are provided in the Harness Grounding paragraph of Appendix K. For EvoHarness-RL, consistent with other training-based baselines, the Experience bank is frozen at the selected training checkpoint. Evaluation episodes can retrieve from this fixed bank, but no consolidation or persistent updates occur. For EvoHarness-Base, consistent with inference-time self-evolving baselines, the Experience bank starts empty and evolves across evaluation episodes using only information from previously completed episodes, without access to future tasks. Consequently, earlier evaluation episodes can inform later ones through accumulated experience, while model parameters remain frozen throughout.

Revision: We added a concise wall-clock training efficiency comparison in Appendix B.2, accounting for harness-related runtime overhead.


Happy to have further discussion!

Thank you again for your thoughtful review and constructive feedback. We have carefully addressed your questions and incorporated the necessary revisions to improve the clarity of our paper. We hope our responses have addressed your concerns, and we would be happy to discuss any further questions or suggestions!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We sincerely appreciate the time and effort that the ACs and reviewers have devoted to evaluating our submission. As the rebuttal discussion period approaches its end, we provide this brief summary in the hope that it helps with your assessment.


Main Strengths Acknowledged by the Reviewers

To begin with, we thank all reviewers for recognizing the main strengths of our work:

  1. A meaningful and well-motivated research problem, investigating how agent policies can effectively coordinate with external runtime harnesses, addressing the challenges of manually designed and controlled harness support across heterogeneous agentic environments. (Reviewers X7qy, ujQm)

  2. A unified BPE abstraction for learning harness coordination, organizing heterogeneous external support through a shared policy-facing interface, where agents learn when and how to selectively interact with their runtime harnesses through cost-aware reinforcement learning. (Reviewers X7qy, jhgw, ujQm)

  3. Strong empirical performance, demonstrating substantial improvements in task success rates and RL efficiency across three diverse agentic benchmarks. (Reviewers X7qy, jhgw, ujQm)

  4. New insights into harness annealing and cross-environment generalization, supported by extensive ablations and analyses revealing how learned harness coordination evolves during training across environments. (Reviewers X7qy, jhgw, ujQm)


How We Addressed the Reviewers’ Concerns

During the rebuttal phase, we made substantial efforts to address the reviewers' concerns through additional experiments, implementation analysis, and methodological clarifications. Below, we summarize the key concerns and how we addressed them:

  1. Clarifying the scope of harness unification and resolving misunderstandings about the framework.

    We clarified that BPE unifies the policy-facing harness coordination interface, grounded in environment-specific support modules and distinct from native environment actions in Aenv\mathcal{A}_{\mathrm{env}}. This unification turns harness coordination into a shared, learnable decision problem, enabling systematic investigation of the complementary roles and evolution dynamics of agent policies and harnesses.

  2. Clarifying the purpose and engineering effort of low-level harness instantiation.

    We explained how environment adapters enable the integration of off-the-shelf, domain-specific harness modules and enhance the framework's adaptability across diverse tasks and environments. Our analysis further shows 65.7–75.9% core harness code reuse, suggesting that relatively little additional implementation effort is required.

  3. Extending generalization experiments to challenging coding tasks.

    On SWE-bench Verified, EvoHarness-Base improves a closed-source frontier model by 4.4 percentage points, while EvoHarness-RL outperforms the GRPO result by 5.0 points with Qwen3-8B.

  4. Strengthening efficiency analysis and clarifying evaluation protocols.

    Additional wall-clock analysis demonstrates faster RL training even after accounting for harness overhead. We also clarified Experience bank isolation for EvoHarness-RL and controlled cross-episode evolution for EvoHarness-Base.

Based on the reviewers' feedback, we have also incorporated necessary revisions into the manuscript, including the Figure 1 caption, Section 2.2, and Appendix B.2, to improve the clarity of our framework.


We sincerely appreciate the reviewers' constructive feedback. We hope our responses address your concerns, and we are happy to discuss if you have any further questions!

Thanks and regards,

#26876 Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.