acceptodds
Under review as a conference paper at ICLR 2027

OpenAgency: A Self-Evolving Multi-Agent System via Governed Collaboration Protocols

Abstract

Large language model (LLM) agents have demonstrated strong capabilities in solving complex tasks, motivating the development of multi-agent systems (MAS) that coordinate agents with specialized roles. However, existing multi-agent collaboration paradigms are often static, with collaboration patterns determined a priori and remaining unchanged during execution. Enabling continuously improving collaboration requires protocols that balance flexibility and controllability: flexibility allows agents to explore diverse coordination strategies at runtime, while controllability ensures that such decisions can be used to improve collaboration. Existing collaboration protocols typically fall into two extremes: they are either rigidly predefined, limiting adaptation, or fully free-form, sacrificing oversight. In this paper, we introduce OpenAgency, a self-evolving multi-agent system built upon governed collaboration protocols. These protocols enable agents to dynamically select their next collaborators within a predefined yet extensible space at runtime, transforming the collaboration pattern itself into an evolvable object. To evolve collaboration protocols, OpenAgency maintains execution traces from agent interactions and introduces a collaboration-driven evolving agent that distills these experiences into a bounded number of atomic protocol improvements. Through this mechanism, OpenAgency allows multi-agent systems to continuously refine their collaboration strategies and adapt to diverse complex tasks while preserving control over the evolution process.Experiments across fourteen benchmarks spanning general question answering, long-horizon reasoning tasks, and production-oriented workloads demonstrate that OpenAgency achieves improved collaboration performance and continues to enhance its effectiveness through self-evolution.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent System

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_m6oROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes OpenAgency, a multi-agent system that treats collaboration protocols as main objects for governance and evolution, aiming to strike a balance between flexibility and controllability. Authors demonstrate the effectiveness of the proposed approach through experiments across 14 benchmarks.

Critical strengths
  1. The paper has a relatively complete structure and content.

  2. The experiments covered a wide range of configurations.

Critical weaknesses
  1. The paper does not sufficiently explain how the initial protocol was constructed and how the experimental baselines were configured. In particular, the construction of the initial protocol (e.g., initial agent number, role assignment, collaboration topology, and completion conditions) and the detailed configurations of the single-agent and multi-agent (init) setups are unclear.

  2. The paper does not sufficiently demonstrate that the optimization effect comes from trace-driven evolution. Since the initial protocol design is not clearly specified and may contain task-specific limitations, the observed gains could partly come from refining an initially inadequate collaboration structure. More analysis is needed to evaluate the robustness of the evolution process under different initial protocols.

Questions for authors
  1. How is the initial collaboration protocol constructed? Specifically, what are the initial agents, role assignments, collaboration topology, and completion conditions?

  2. How sensitive is the evolution process to different initial protocols?

Reviewer_jq3COpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OpenAgency is a multi-agent system that evolves by revising its collaboration protocols. A protocol defines roles, collaboration nodes, admissible handoffs, handoff modes, and completion conditions. During execution, agents can select one, some, or all admissible downstream nodes. The governance layer checks handoff validity, completion conditions, and predecessor dependencies before activating downstream execution.

The system records accepted and rejected handoffs, completion events, and failures. An evolving agent analyzes these traces and proposes up to four edits per round, modifying the protocol's structure, conditions, or shared rules. Candidate protocols first undergo structural validity checks, after which their adoption is determined by the collaboration quality score assigned by the evolving agent.

The experiments use GPT-5.5 and conduct four rounds of protocol evolution separately on fourteen benchmarks, comparing a single agent, an initialized multi-agent system, evolved protocols, and other frameworks. The authors report that the final system outperforms all listed comparison conditions on each benchmark's chosen primary metric, with an average gain of 20.7 percentage points over the initialized multi-agent system. The paper also provides search-space ablations, protocol evolution examples, analyses of rejection and recovery behavior at the governance layer, and execution-process metrics for software tasks.

Critical strengths
  1. The distinction between the protocol specifying who may receive a handoff and the executor deciding who actually receives it in a particular run is useful. The governance layer checks target nodes, completion conditions, and predecessor dependencies, giving collaboration requirements concrete runtime enforcement beyond relying on roles to follow their prompts. Rejected handoffs are also logged, making it easier to inspect executions and identify rules that may need revision. I consider this one of the paper's most practically valuable aspects.

  2. The software repair analysis identifies a concrete improvement: the proportion of instances producing an applicable patch rises from 79.2% to 100%, while the file-match rate conditional on having a patch changes relatively little. This suggests that reducing execution failures that prevent patch delivery is an important source of the gain. The protocol evolution examples also include adding verification and retries and removing unnecessary roles, showing that improvement does not rely solely on increasing system size. The paper further reports cases in which execution recovers after a rejected handoff. These analyses help clarify which execution problems the system addresses.

Critical weaknesses
  1. The paper conflates a predefined topology with a predefined execution path. Even with a fixed graph, execution can select branches, retry, or skip a role based on the current state. AutoGen's FSM Group Chat already constrains role communication through a transition graph while retaining conditional selection and loops. AFlow's code-based workflows can likewise express conditions and loops and be revised using execution feedback. Flexible runtime choices and structural revision in subsequent rounds are therefore not capabilities that prior methods were unable to combine.

    OpenAgency's more specific feature is its integration of completion conditions, rejection events, and protocol revision within a unified interface. This has engineering value, but I see it as an extension of existing constrained workflows and automated optimization methods. The evolution procedure still consists of model-based diagnosis, local edit proposals, and greedy candidate acceptance, which does not change my assessment of the contribution. I do not think these differences are sufficient to substantiate the claimed new collaboration paradigm.

  2. Runtime validation, candidate acceptance, and actual task quality need to be discussed separately. The governance layer checks whether behavior conforms to the current protocol, while evolution checks whether a candidate is structurally valid and whether its model-assigned collaboration score decreases. Neither check directly guarantees that task quality will not degrade. Completion conditions and role requirements can themselves be revised, so satisfying a new protocol does not establish improvement while preserving the original task requirements.

    The experimental attribution is also too strong. Compared with a single agent, the full system simultaneously changes the division of responsibilities, context organization, and handoff mechanisms. This comparison cannot determine whether the gains come from governance or from other aspects of the multi-agent design. On OfficeQA Pro, the condition labeled MetaGPT permits only revisions to responsibilities within a single node, whereas the full method can add nodes and connections. This result primarily supports the usefulness of a broader structural search space; it does not directly establish that governed protocols are superior to SOP-based collaboration in general.

    The reported rejection and recovery behavior supports the usefulness of runtime constraints in some instances, but it cannot replace controlled comparisons of the mechanisms above. The current results more directly show that automatically adjusting roles, verification, and recovery procedures can be beneficial. They do not establish that governance provides collaboration evolution with the control guarantees emphasized in the paper.

Questions for authors
  1. What specific capabilities does OpenAgency add beyond existing constrained transition graphs and workflows with conditions and loops? Does it change the expressive capabilities or execution semantics of the protocol, or does it make existing capabilities easier to record and revise? Please use an actual evolved protocol to distinguish improvements that depend on specific governance mechanisms from those that could also be achieved through ordinary changes to roles, prompts, or connections.

  2. When completion and transition conditions can be revised, which task requirements must remain invariant? How does the system distinguish these requirements from internal collaboration rules that may be adjusted? A rejected handoff could reflect either an incorrect request by the executor or an overly restrictive protocol. What evidence does the evolving agent use to decide whether to correct the behavior or revise the rule? Is collaboration quality before and after revision evaluated against the same criteria, independent of changes to the protocol? Does "non-degradation" refer specifically to a non-decreasing observed proxy score, or does it also include task constraints that must always remain satisfied?

  3. Is the single-node condition labeled MetaGPT on OfficeQA Pro a reproduction of the framework or an ablation that restricts structural search? These settings support different conclusions. In the governance analysis, how is the recovery rate attributed to rejections that "changed the outcome" determined? Have you conducted a controlled comparison that holds all other conditions fixed, allows an originally rejected handoff to proceed, and checks whether the execution outcome changes?

  4. Which tasks were used to collect evolution traces, evaluate candidate protocols, and report final performance? Are these task sets independent? If the final evaluation tasks were repeatedly used during evolution, the results should be interpreted as adaptation to those instances, rather than direct evidence of improvement on unseen tasks. Please clarify the actual setup and whether the comparison conditions had the same opportunity to adapt. For software repair, please also report the actual issue-resolution rates of the main comparison conditions under the same setup, so that improvements in file localization, patch delivery, and successful repair can be distinguished.

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

This paper introduces OpenAgency, a self-evolving multi-agent system built on governed collaboration protocols. The protocols explicitly define roles, permitted handoffs, and completion conditions, allowing agents to choose collaboration paths at runtime within these constraints. Execution traces are then used to iteratively revise the protocol structure, conditions, and shared rules. The authors evaluate the method on fourteen benchmarks and report improvements over the initial system and the selected baselines.

Critical strengths

• The system design has merit. It combines runtime collaboration constraints with protocol optimization across runs, providing an explicit framework for managing agent coordination. • The paper is well organized. The architecture diagram and protocol evolution examples help explain the method. • The experiments cover several task types, and the reported results suggest potential practical value.

Critical weaknesses

• The novelty is not sufficiently established. Dynamic collaboration, workflow optimization using execution feedback, and structural search have been explored in prior work. The potential contribution here lies in using governance events to guide protocol revisions, but the paper does not clearly establish the added value of this mechanism. • The paper discusses automated optimization methods such as AFlow and ADAS but does not compare against them under comparable search budgets. The current experiments therefore do not establish whether the proposed evolution mechanism improves on existing optimization methods. • Computational costs and baseline configurations are insufficiently reported. The paper provides neither token usage nor model-call costs for search and deployment. A fixed tool budget does not establish comparable total compute, and it is unclear whether the single-agent and other baselines receive comparable opportunities for optimization. • Several central mechanisms lack implementation details. Trace diagnosis, state-dependent validity checks, completion conditions, and candidate quality scoring are described mainly through abstract functions, without concrete rules, prompts, or worked examples. • The ablations do not adequately isolate the contributions of the core components. They mainly restrict the structural search space, rather than directly testing rejected-event feedback, runtime constraints, or candidate quality filtering. They also do not sufficiently separate the gains from structural changes from those due to shared-instruction optimization.

Questions for authors
  1. Section 2 discusses automated optimization methods such as AFlow and ADAS. Could the authors compare against at least one of these methods under comparable search budgets, executor models, and feedback conditions to assess the advantage of the proposed evolution mechanism?
  2. Could the authors report both the total token usage and computational cost of the four evolution rounds and the per-task inference cost of the final protocol? Search costs should include trace diagnosis, edit generation, candidate execution, and scoring. Do the single-agent and other baselines receive comparable total compute budgets and opportunities for optimization?
  3. How are the state-dependent validity checks, completion predicates Ωn, diagnosis function fdiag, and collaboration score qcol in Sections 3.2–3.3 implemented? Which rely on deterministic rules, and which rely on LLM judgments? Please provide a complete example covering the execution trace, diagnosis, proposed edits, and candidate acceptance decision.
  4. Could the authors separately ablate rejected/failed-event feedback, runtime governance, and candidate quality filtering? It would also help to distinguish the gains from structural modifications from those due to shared-instruction optimization.

Comments on the paper

2 comments

Dear Reviewers,

Could you please review the authors’ rebuttals and follow up with any further comments or questions?

Many thanks for your time and efforts.

Best regards, AC

Dear Reviewers,

As the AC–reviewer discussion phase is about to begin, please highlight any disagreements or inconsistencies across the reviews, as well as any concerns that remain unresolved after the author rebuttal.

Please share your perspectives on these points and engage with one another to help us reach a well-supported assessment of the paper.

Thank you for your time and contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.