acceptodds
Under review as a conference paper at ICLR 2027

Ouroboros: A Self-Developing Coding Harness with Reviewed Core Evolution

Abstract

Long-horizon agents are model–harness systems, yet most harnesses remain fixed after design. We present Ouroboros — a self-developing agent harness whose tools, context assembly, prompts and core implementation improve through reviewed commits that become the runtime for later work. Core evolution proceeds in two modes. In recursive free evolution, improvement is itself a task and completion can schedule the next evolution cycle. In experience-driven core evolution, ordinary work and social interaction expose bugs, rough edges, and inefficient context construction leading to reviewed structural changes. On Terminal-Bench 2.1, an Opus 5 run scores **86.97%** (86.74% after trajectory audit), the best result reported on this benchmark. An Opus 5 run on OSWorld-Verified reaches **90.69%**, above the best previously reported score, and a five-rollout CL-Bench campaign sets a new state of the art at **0.2301** **(all three comparisons as of August 2026)**. Hope is the longest-running publicly documented Ouroboros deployment: a 161-day living-agent experiment in free evolution under governed human communication across seven surfaces, where people surface faults and proposals but the agent decides which changes to pursue. Because a self-developing agent may rewrite its own code and select new model APIs, operational safety is a primary design problem: guardrails must remain authoritative under evolutionary pressure. Benchmark campaigns use frozen seeds, while Hope continues live evolution on a separate lineage.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agent, LLM, self-evolving, subagents, NLP

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gnZxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents Ouroboros, a self-developing agent harness. The harness operates in two modes: recursive free evolution, where the harness specifically attempts to evolve itself, and experience-driven core evolution, where the harness performs regular tasks. The authors present the architecture of the system. The system is evaluated on 3 benchmarks: TerminalBench 2.1, OSWorld-Verified, and CL-Bench and attains state-of-the-art performance on all three. The authors also provide a qualitative report on Hope, a deployment of Ouroboros that has been running for 210 days at the time of submission. Finally, the authors discuss guardrails they use to maintain operational safety of the self-evolving system.

Critical strengths
  • The authors study a timely and relevant problem: self-improving agent harnesses
  • The qualitative study on Hope, an extremely long-horizon AI system, is interesting and novel
Critical weaknesses
  • Presentation. The paper is in general very difficult to follow.

    • Table 1 presents several items that are either present or not in various self-evolving systems. However, these items are (a) not of the same type (being able to edit prompts and having every change implemented as a "reviewed commit" are two very different features of a system), (b) insufficiently described. For example, how is a workflow distinct from core code? What is deployment state in this context?
    • Figure 1 introduces many different concepts which are never discussed in the figure or the text. Take, for example, 'Typed Outcomes'. There is no description of this component and it is left to the reader to imagine what it entails or why it's important. It is also unclear what the arrows in the figure are meant to denote.
    • Several sections of the text introduce concepts without fully describing them. Example: "The runtime represents identity and continuity through a versioned constitution, an editable identity profile, scratchpad and chronicle projections, project memory, review ledgers, and Git history. ". It is generally hard for a reader to understand what these components are and why they are necessary.
    • The two modes of core evolution are unclear. Arguably, the two-way decomposition of the evolution process is the primary distinguishing element of Ouroboros. The authors should make the distinction clearer, potentially with a symbolic description or examples. Is the difference purely whether evolution is a primary of secondary goal of the task?
  • Missing details on Hope. The description of Hope is lacking in several ways.

    • The authors state that Hope has been running since 27 February, 2026. However, time is arguably a poor metric for measuring the lifecycle of an asynchronous agent. The number of tasks and evolution cycles would be better metrics to report.
    • The authors describe many features of Hope (Multi-channel state, Controls on self-directed work), but it is not clear whether these are common to all Ouroboros deployments. If these are common, these would be better placed in the architecture description.
  • Evaluation protocol is entirely missing. Section 5 describes a number of results, but these are largely meaningless without descriptions of the evaluation settings, which are missing. Were these run with Hope or without Hope? At what stage of the lifecycle were these experiments run? What data did the agent have access to in order to evolve?

  • Missing baselines. The authors do not benchmark against other self-evolving systems.

  • Missing architectural ablations. The authors present several aspects of the Ouroboros architecture that are not adequately justified or ablated. While many of these features are interesting, based on the paper I am not sure why they are necessary. Examples include features described in Figure 5, sub-agents, and the 3-tiered commit pipeline.

  • No justification for choice of safety controls. The authors do not provide any concrete quantitative evidence that their "Operational Safety Controls" are effective. An ablations demonstrating misaligned behavior would be good evidence that these measures are justified.

Questions for authors
  1. Can you provide an example of a task from each of the two evolution phases?
  2. How does the agent decide when to terminate each phase?
  3. What Ouroboros system are the evaluation results in section 5 for? i.e. what evolution process did it undergo?
  4. What data was the Ouroboros system evaluated in section 5 developed on? How did you ensure there was no leakage of test set data?
  5. What is the minimal set of Ouroboros components needed to achieve the evaluation results in section 5?
Reviewer_iJy6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a self-evolving agent harness named Ouroboros, with two types of evolution mechanisms. Ouroboros can either treat self-evolution as a task or learn from general interactions and human feedback and independently decide whether to evolve. Changes to the Ouroboros code go through an auditing mechanism. Ouroboros also reports state-of-the-art results on several popular benchmarks. The paper also introduces Hope as a long-running self-evolving agent that has been running for at least 210 days.

Critical strengths

The paper proposes a self-evolving agent framework, and its continuous deployment and evolution are particularly interesting.

Critical weaknesses
  1. The presentation could be clearer. The paper introduces many technical terms without sufficiently explaining what they refer to or how they are implemented, such as "Diff-review Panel," "Identity-edit," and "One-writer Lease," among others. In addition, the descriptions of the proposed agent framework's mechanisms and behaviors are somewhat vague, making it challenging to assess the specific technical contributions and challenges addressed by the work.
  2. The evaluation section mainly reports benchmark scores without much further analysis. It is difficult to understand how the self-evolution process changes the agent's behavior over time, and the paper would benefit from more in-depth analysis and insights into these changes.
  3. Although the paper presents Hope as a long-running self-evolving agent, it does not provide a systematic evaluation of Hope itself, making it difficult to assess the effectiveness of its long-term self-evolution.
  4. It is unclear how the Ouroboros harness used for benchmark evaluation was developed through self-evolution, and whether the authors repeatedly evolved and evaluated it against the same benchmarks. The authors could provide a more systematic analysis of potential benchmark information leakage during self-evolution and discuss mechanisms to prevent it. It would also be valuable to investigate whether improvements in benchmark scores reflect overfitting to specific benchmarks or more generalizable improvements in the agent's capabilities.
Questions for authors
  1. How was the Ouroboros harness used for benchmark evaluation developed? Was it repeatedly evolved on the same benchmarks? How did the authors prevent benchmark information leakage and distinguish benchmark-specific overfitting from generalizable improvements?

  2. Could the authors provide more analysis of how self-evolution changes the agent's behavior and capabilities over time, beyond the reported benchmark scores and particularly in the long-running Hope deployment?

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

Ouroboros is an agent harness that modifies its implementation through reviewed commits, using either dedicated improvement tasks or failures encountered during ordinary work. The paper evaluates frozen harness versions across five benchmark families and describes Hope, a 210-day deployment with persistent memory, human interaction, and continuing self-modification.

Critical strengths

Broad evaluation and careful auditing. The paper reports competitive performance across several demanding benchmarks, including comparisons across models. Correcting a rewarded Terminal-Bench shortcut, retaining infrastructure failures in the denominator, and symmetrically filtering reference leakage in SWE-bench Pro are valuable methodological practices.

Substantial deployment experience. The long-running deployment, concrete repair examples, and explicit accounting definitions provide useful evidence that persistent self-modification is operationally feasible. The reviewed commit pipeline and recovery mechanisms address practical problems beyond isolated benchmark execution.

Critical weaknesses

The benefit of core evolution is not established. The benchmark campaigns evaluate frozen seeds, with evolution explicitly disabled for four benchmarks in Table 3, while Hope follows a separate lineage. This measures the final harness, but the manuscript does not connect its performance to particular autonomous improvements. The deployment offers two repair examples and activity counts without measured before/after outcomes. CL-Bench also requires a distinction between performance and learning: the public artifacts report higher normalized reward, but normalized gain of 0.1985 versus 0.2406 for same-model ICL. Since gain measures improvement relative to each system’s stateless baseline, the reward result alone does not establish stronger continual learning.

The distinctive contribution of reviewed evolution needs stronger specification. DGM and SICA already modify agent implementations; SICA also includes LLM oversight. Autogenesis specifies evaluated, versioned updates with conditional commit and rollback. Ouroboros’s particular deployment and review architecture is therefore the most relevant distinction. However, section 3 and 7 and Appendix A do not adequately specify reviewer composition, quorum, or which enforcement components remain outside the agent’s modification authority when review logic itself is mutable. Blocked-commit counts and observed shutdown compliance do not establish that accepted changes preserve these boundaries.

Questions for authors

Which autonomous changes distinguish the evaluated frozen harness from its starting version? Can existing logs or a small comparison under the same model and budget show their effect on task performance or the reported repair outcomes?

Which components independently enforce approval before modified code becomes active, particularly when review logic changes? Please specify the reviewer models, quorum, and acceptance criteria.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.