acceptodds
Under review as a conference paper at ICLR 2027

Autogenesis: A Self-Evolving Agent Protocol

Abstract

LLM-based agent systems can adapt by revising prompts, constructing tools, and changing other runtime resources. Managing these heterogeneous updates requires a shared lifecycle for identifying targets, assessing candidates, and retaining accepted versions, complementing communication protocols such as A2A and MCP. We introduce AUTOGENESIS PROTOCOL (AGP), which decouples what evolves from how updates are proposed. Its Resource Substrate Protocol Layer (RSPL) represents prompts, agents, tools, environments, and memory as protocol-registered resources¹ with explicit state and versioned interfaces. Its Self-Evolution Protocol Layer (SEPL) specifies candidate proposal, assessment, and commit, with auditable histories and support for restoring retained resource states. We instantiate AGP in AUTOGENESIS SYSTEM (AGS), a self-evolving multi-agent system that retrieves, instantiates, and refines resources during execution. Evaluations characterize task-level performance across reasoning, tool use, and coding, with output-token costs reported for repeated reasoning runs. Execution cases illustrate a shared lifecycle across heterogeneous resource types.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM-based Agents, Self-Evolution, Agent Protocol, Multi-Agent Systems, Resource Lifecycle Management

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_eWs7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Autogenesis Protocol (AGP), a two-layer framework for managing updates to heterogeneous resources in LLM-based agent systems. The authors instantiate AGP in Autogenesis System (AGS), a planner–specialist multi-agent system with reflection-driven updates. The paper positions AGP as a shared resource-update lifecycle rather than a new model-training algorithm or a replacement for communication protocols.

Critical strengths
  • Relevant systems problem.
  • Clear conceptual distinctions.
Critical weaknesses
  • The experiments do not isolate the protocol’s contribution from the optimizer, additional inference, and task-local refinement.
  • Several central protocol properties remain interface requirements or partial implementation descriptions rather than demonstrated capabilities.
  • State-reset ambiguity, limited statistical analysis, and incomplete coding-result accounting weaken reproducibility.
Questions for authors
  1. The main contribution is a shared lifecycle for heterogeneous resource updates, whereas the quantitative experiments primarily compare configurations with and without additional refinement. These results do not isolate the contribution of the proposed protocol.
  2. Definitions 3.1–3.5 formalize resource management through typed records and composable operators, yet Appendix G.2.3 reveals heterogeneous Python interfaces with several operations requiring additional adapters. This raises questions about protocol completeness and independent AGP-compliant implementations.
  3. The Reflect–Select mechanism assumes that execution failures can be attributed to specific editable resources, but the paper does not evaluate the accuracy of resource selection or the consequences of incorrect attribution.
  4. The framework supports heterogeneous resource updates, but dependencies among prompts, agents, tools, and environments may require coordinated modifications. How does AGP handle cross-resource dependencies, conflicting updates, and consistency during joint evolution?
  5. The acceptance gate is central to SEPL, yet its reliability is not evaluated. How frequently do self-assessed improvements correspond to actual downstream gains, and what are the false acceptance and rejection rates?
  6. Although AGP aims to decouple resource evolution from update strategies, all experiments rely on reflection, while alternative adapters remain unevaluated and sometimes omit the required acceptance gate. Thus, the claimed optimizer independence is not empirically established.
  7. Appendix D.1 states that benchmark resets do not necessarily reset resource registries, memory stores, or caches. Please clarify whether tasks, methods, and repeated runs start from identical resource states and whether updates persist across questions.
  8. The reasoning results demonstrate a higher-cost operating point rather than an isolated advantage of resource evolution. Comparisons against standard refinement or search methods under matched total inference budgets would strengthen the conclusions.
  9. The protocol emphasizes versioned updates and restoration, but no experiments evaluate recovery after faulty activation or interrupted commits. Controlled failure-injection tests would better establish the reliability of its lifecycle guarantees.
Reviewer_SFeKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Autogenesis Protocol, a framework that separates evolution from proposing new updates. Within the framework, two core components are (1) Resource Substrate Protocol Layer (RSPL) manages the evolution of resources with recorded versions, and (2) Self Evolution Protocol Layer (SEPL) proposes new candidates, evaluates and then commits, the idea itself is novel to me. The author’s experiments demonstrate substantial improvements. But I have important concerns about the experiments and baselines.

Critical strengths

(1) The idea of separating resource evolution from harness updates is novel. These asynchronous updates could be very useful in many high-stakes domains where resource updates take substantially longer time or require additional efforts. The decoupling is good for improving the overall efficiency and framework.

(2) Empirical experiment results demonstrate clear improvements on GPQA-Diamond, GAIA, AMIE 24 and 25 benchmarks. The authors reported Prompt-Evo, Solution-Evo and joint evolutions of both; these scores provide detailed information of how evolving different resources contribute to the performance gains.The joint optimization shows consistent improvements compared to optimizing either of the prompt or solution.

(3) On code agent benchmark, the proposed method again shows clear improvements compared to baseline models. The Solution-Evo shows at least 10% gains.

Critical weaknesses

(1) There are important baselines missing in the paper, preventing the work from publication. First, many baselines in this work are single LLMs while the proposed method is a multi-agent system. Therefore, it is unfair and more appropriate baselines would be top LLM harnesses, including Claude Code, CODEX, and other open-source harnesses.

(2) While the authors try to prove the efficacy of the Autogenesis Protocol, the comparison with other RSI frameworks is necessary. I would suggest the authors consider baselines including RRSI, Dream-RSI, etc.

(3) The benchmarks in this paper are limited, GPQA-Diamond, GAIA, AMIE 24 and 25 are standard benchmarks and exhaustively studied in numerous papers. To justify the value of the proposed method, the authors should run evaluations on more challenging benchmarks, such as Terminal-Bench 2.0 and SWE-bench Pro, which would better assess adaptation during complex, long-horizon execution.

Questions for authors

(1) Since the resource update and execution update are separated, I’m wondering if the RSPL can see the evolution history of SEPL, and if SEPL can see the evolution history of RSPL. Some information in history can provide incredibly valuable hints for future iterations. The author should clearly state how the history information is being used under current decoupling.

(2) Is the resource a fixed set? Does the current framework support introducing newly emerged tools into the current resource pool? The authors should clearly discuss whether current resource updates are only applied to existing resources.

(3) Some numbers in the table look suspicious. In Table 1 GPT-4o and GPT-4.1 sections, Prompt-Evo has the exact same number as Solution-Evo. In Table 2 GAIA, AWorld has the exact same number as Langfun.I would recommend the authors double check the table and make sure all numbers are correct.

Reviewer_1ZoUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose AGP (AUTOGENESIS PROTOCOL), a two-layer protocol. 1) RSPL (Resource Substrate Protocol Layer) includes prompts, agents, tools, environments, and memory as versioned resources. 2) SEPL (Self-Evolution Protocol Layer) defines a lifecycle of propose → evaluate → commit over those resources. AGS (AUTOGENESIS SYSTEM), a planner–specialist multi-agent system, instantiates it. Results on GPQA-Diamond/AIME, GAIA, HLE show improvement, as well as on a 100-problem LeetCode set, plus three qualitative cases covering all five resource types.

Critical strengths
  1. The core idea of version control for an agent is useful and can be important as agent harnesses become more fundamental for complex agentic workflows.
  2. The experiments cover a range of agents and models.
  3. The paper's presentation of these ideas is systematic and rigorous.
  4. The results show clear improvement using the method.
Critical weaknesses
  1. The paper is lacking in background, referring to other methods that aim to standardize agent workflow, like Harbor, Exgentic, and Cube. Surveys and position papers that point to this area of research. I think these can enhance and emphasize the importance and contribution of the work.
  2. The paper's experiments, although extensive, raise several questions: why general coding agents like Codex, Claude Code, and Pi are not assessed (at least the open ones). The benchmarks tend towards LLM-focused ones like AIME, GPQA, and HLE, while Gaia is already a bit older and saturated. No comparison under a fixed compute budget.
  3. The results presentation makes it hard to understand which underlying model is running each agent and what the main comparison point is.
  4. No support or analysis showcasing the value of the protocol itself directly.
Questions for authors
  1. What does AGS score on HLE without evolution, under the same judge?
  2. Is there any experiment showing a benefit unique to the protocol? Does it recover from a bad commit or a cross-resource coordination failure in a way that plain reflection doesn't?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.