PaperlayerView exact revision on arXiv

Source-linked AI summary

A Responsibility-Centered Coordination Mechanism for Decentralized Multi-Agent Systems

Thi-Nhung Nguyen, Viet-Thanh Pham, Linhao Luo, Thuy-Trang Vu, Dinh Phung

arXiv:2610.21243v1cs.MAcs.AI

TL;DR

Decentralized coordination can leave responsibility for delegated dependencies and recovery unclear, while centralized orchestration has scalability, adaptability, and performance bottlenecks. RCM couples delegation with explicit responsibilities and local commitment management. Across reasoning, coding, and executable environments, it reports gains over centralized and decentralized baselines, including up to 18.7 points in task performance and 10.2 points in overall responsibility.

  • Problem

    Decentralized approaches distribute coordination but may leave responsibility for delegated dependencies and failure recovery unclear.

    Source: §1 · p. 1

  • Method

    RCM couples task delegation with explicit responsibilities and uses recursive delegation, responsibility-constrained decisions, and RCP to manage commitments locally.

    Source: §1 · p. 2

  • Results

    Across reasoning, coding, and executable environments, RCM gains up to 18.7 points in task performance and 10.2 points in overall responsibility over strong centralized and decentralized baselines.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Explicit responsibility for each commitment enables robust and autonomous coordination without sacrificing executor reasoning and coding capability.

    Source: §4.1 · p. 6

    The evaluation does not represent every task, model, or deployment environment for multi-agent coordination.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Centralized orchestration provides comprehensive multi-agent coordination but creates bottlenecks in scalability, adaptability, and collective performance. Decentralized approaches alleviate these bottlenecks by distributing coordination across agents, yet often leave responsibility for delegated dependencies and failure recovery unclear. To address this gap, we propose RCM, a responsibility-centered coordination mechanism that leverages commitments to couple task delegation with explicit responsibilities of the involved parties. Through recursive delegation, agents retain their upstream responsibilities while locally orchestrating downstream commitments, thereby preserving accountability across delegation chains, creating a network of local orchestration. We operationalize this mechanism locally through responsibility-constrained decision-making and a responsibility-oriented communication protocol (RCP), enabling agents to autonomously establish, manage, and terminate commitments. As agents act in parallel, their local actions continuously evolve the commitment graph, allowing global coordination to emerge and adapt toward a shared mission without centralized control. Experiments across reasoning, coding, and executable agent environments demonstrate that RCM outperforms strong centralized and decentralized coordination baselines by up to 18.7 points in task performance and 10.2 points in Responsibility, while improving failure and harm detection without task-specific training.

1 Introduction

The introduction identifies unclear responsibility for delegated dependencies and failure recovery as a gap in decentralized coordination. RCM couples task delegation with explicit responsibilities, enabling local orchestration across commitment chains and global coordination through autonomous decisions.

1 Introduction

  • Centralized orchestration faces scalability, adaptability, and collective-performance bottlenecks, while decentralized delegation can leave ownership of downstream dependencies and recovery unclear.Centralized systems manage dependencies and revise plans after failures, but their coordination overhead and context grow with scale.Source: §1 · p. 1
  • RCM treats commitments as coordination units: the delegatee is responsible for fulfilling the task, while the delegator manages the resulting dependency until termination.Recursive delegation preserves upstream commitments while agents locally orchestrate downstream dependencies.Source: §1 · pp. 1–2
  • Responsibility-constrained decisions and RCP let agents establish, manage, and terminate commitments as local actions evolve the commitment graph toward a shared mission.Agents act in parallel and can respond to local changes, including switching delegatees.Source: §1 · p. 2, Figure 1 · p. 2
  • Across reasoning, coding, and executable environments, RCM gains up to 18.7 points in task performance and 10.2 points in overall responsibility versus strong centralized and decentralized baselines, while improving failure and harmful-trajectory detection without task-specific training.The proposed responsibility-oriented metrics include Communication, Monitoring, Intervention, and Fulfillment.Source: §1 · p. 2

2 Responsibility-centered Coordination Mechanism (RCM)

RCM coordinates agents through linked commitments that pair delegated outcomes with explicit responsibilities. Recursive local orchestration lets parallel local decisions evolve a commitment graph toward a shared mission without a central execution plan.

2.1 Global Coordination through Commitments

  • A commitment tracks the delegation, lifecycle state, and state-dependent responsibilities; the delegator orchestrates until termination while the delegatee fulfils agreed work after acceptance.Fulfilment is bounded by agreed scope and authority, and revisions require renewed agreement.Source: §2.1 · p. 3
  • Recursive delegation links downstream commitments into a dynamic graph, whose local evolution coordinates work without any agent prescribing or observing a global plan.When a downstream commitment is violated, its delegator may replace it while unaffected work continues independently.Source: §2.1 · p. 3, §2.2 · p. 5

2.2 From Local Actions to Global Coordination

  • Agents decide from local views of the evolving state and active responsibilities; the Dec-POMDP formulation is descriptive, not a training procedure.Local actions include fulfilling upstream work, delegating supporting work, and monitoring, verifying, revising, or recovering downstream work.Source: §2.2 · p. 4
  • RCP standardizes negotiation, information exchange, execution reporting, outcome verification, and guarded commitment updates between relevant counterparts.Messages and execution evidence inform decisions and can trigger state transitions only after conditions and evidence are verified.Source: §2.2 · pp. 4–5

3 Experimental Setup

The setup evaluates coordination across mathematical reasoning, code generation, and executable workflows, including controlled failures and changing real-world workflows. Methods share task instances, tool allocations, permissions, prompts, and test-time budgets.

3 Experimental Setup

  • AIME25 tests mathematical reasoning, HUMANEVAL functional code generation, CLAW-EVAL multiservice tasks under controlled failures, and CLAW-EVAL-LIVE changing real-world workflows.The two CLAW-EVAL settings use randomly partitioned tools across agents, requiring delegation and communication.Source: §3 · p. 5
  • All methods use identical task instances, tool allocations, permissions, prompts, and test-time budgets.Source: §3 · p. 5
  • Algorithm 1 begins coordination in parallel from agreed initial commitments and recursively coordinates each delegatee.Source: §3 · p. 5

Appendix A.3 gives benchmark details.

The benchmark details define task-capability and responsibility measures, identify comparison baselines, and describe how Table 2 encodes coordination support.

Appendix A.3 gives benchmark details.

  • Responsibility measures failure handling through Communicate, Monitor, Intervene, and Fulfillment; their equal-weight mean forms the Responsibility score.Task capability uses accuracy on AIME25, Pass@1 on HUMANEVAL, score and completion on CLAW-EVAL, and score on CLAW-EVAL-LIVE.Source: Appendix A.3 · p. 5
  • Baselines span decentralized, hand-designed, and optimized multi-agent methods, with Aegis attribution models and training-free LLMs included on Who&When.ATBench-C comparisons include guard models and general-purpose LLMs.Source: Appendix A.3 · pp. 5–6
  • Table 2 encodes whether workflows are dynamic, how responsibility is specified, and whether distributed monitoring and recovery are supported.Its symbols ✓, ◦, and × denote full, partial, and no support.Source: Table 2 · p. 6
  • Figure 3 concerns agent scaling on CLAW-EVAL.Source: Figure 3 · p. 6

4 Results

The results examine coordination across benchmarks and changing teams, then assess responsibility, safety monitoring, and emergent organization. RCM leads on agentic tasks, scores above the strongest baseline on responsibility, and exhibits adaptive coordination patterns.

4.1 Robustness and Autonomous Adaptation

  • RCM improves CLAW-EVAL by 18.7 points with Qwen3-4B and 6.3 points with GPT-5.4, and CLAW-EVAL-LIVE by 1.9 and 7.5 points, respectively, over the strongest baseline in each column.With Qwen3-4B, RCM achieves the best result in all four reported columns; with GPT-5.4, it leads both agentic benchmarks, ties the best AIME25 result, and remains within 1.5 points of the best HumanEval result.Source: §4.1 · p. 6
  • As teams grow from 1 to 16 agents with tools divided more finely, RCM maintains the strongest task performance and becomes more efficient at larger team sizes.Several baselines are faster with small teams, but their latency grows more sharply as agents are added; RCM does not require a separate workflow for each configuration.Source: §4.1 · p. 6

4.2 How Responsibility Shapes Coordination

  • RCM reaches a Responsibility score of 82.8%, 10.2 points above the strongest baseline, with 100.0% Communicate, 99.0% Monitor, 67.0% Intervene, and 65.4% Fulfillment.GPTSwarm reports 95.2% Communicate and 95.0% Monitor but 32.5% Intervene and 37.8% Fulfillment.Source: §4.2 · pp. 6–7, Table 3 · p. 7
  • On Who&When, RCM with GPT-5.4 achieves an average F1 of 34.96, outperforming the strongest selected baseline, o3, by 9.83 points.Who&When tests identification of the responsible agent and decisive error step.Source: §4.2 · p. 7, Table 4 · p. 7
  • On ATBench-C, RCM with Qwen3-4B-Instruct-2507 achieves an F1 of 73.82, surpassing the strongest selected guard model by 2.72 points without task-specific training.Source: §4.2 · p. 7, Table 5 · p. 7
  • Removing any RCM component degrades performance, and restricting initial task receivers to one agent also performs worse than assigning the task to all agents initially.The tested single-agent assignments use either relevance-based or random selection.Source: §4.2 · p. 7, Table 6 · p. 7
  • RCM trajectories form linear support, hybrid networks, and specialist hubs, while agents dynamically recruit missing expertise and adapt delegation to execution needs.Agents also show low communication breadth but high coordination depth; reports are often environment-grounded, and rejected actions prompt revision while unresolved dependencies prevent premature closure.Source: §4.2 · pp. 7–8, Table 7 · p. 8
  • Adaptive self-organization lets coordination structures, delegation strategies, and expertise roles evolve under stable responsibility constraints, reducing reliance on manually engineered workflows and roles.Source: §4.2 · p. 8

5 Related Work

Prior work studies commitments, protocols, designed workflows, and decentralized LLM-agent coordination. RCM extends commitment-based coordination by making delegators explicitly responsible for delegated dependencies across recursive delegation.

5 Related Work

  • Classical multi-agent coordination formalizes collective activity through joint intentions, social commitments, and protocols governing roles and dependencies.Source: §5 · p. 8
  • Recent LLM-based systems use designed role-based workflows or learned orchestration structures, with dependencies governed by the resulting workflow.Source: §5 · p. 8
  • RCM extends social commitments with explicit delegator responsibility and operationalizes joint responsibility through obligations maintained across recursive delegation.Each delegator acts as a local orchestrator for its delegated work while preserving agent autonomy.Source: §5 · p. 8
  • Decentralized LLM-agent work explores communication, collective reasoning, planning, routing, and inference-time control, but does not explicitly assign responsibility for dependency monitoring, outcome verification, and failure recovery.Source: §5 · p. 8

6 Conclusion

RCM assigns explicit responsibility for task fulfillment and local orchestration, supporting autonomous coordination and recovery without centralized control. The conclusion reports stronger coordination performance, improved safety and resilience, and communication costs closer to linear scaling.

6 Conclusion

  • RCM assigns responsibility for task fulfillment and local orchestration to the involved parties, orienting decisions and communication toward those responsibilities.The conclusion describes this as a mechanism for decentralized multi-agent systems.Source: §6 · p. 8
  • Local orchestration enables agents to detect harmful actions and execution failures, initiate local recovery, and adapt to changes.The conclusion says diverse coordination patterns emerge without prior specification.Source: §6 · p. 8
  • Bounding orchestration and communication to local commitments reduces coordination overhead and brings communication costs closer to linear O(n) scaling.Source: §6 · p. 8

Limitations

The evaluation spans reasoning, coding, and executable agent environments, two executor models, and differently sized teams, but cannot represent every deployment setting.

Limitations

  • The evaluation covers reasoning, coding, and executable agent environments, two executor models, and teams of different sizes.Source: Limitations · p. 9
  • These evaluation settings cannot represent every task, model, or operating environment in which multi-agent coordination may be deployed.Source: Limitations · p. 9
  • The authors identify broader deployment settings and more ambitious problems as directions for future work.Source: Limitations · p. 9

Ethical Considerations

The authors caution that autonomous coordination can reduce direct human oversight and may amplify harmful actions under specified conditions. They also note the substantial computational cost of evaluation and the measures used to limit it.

Ethical Considerations

  • Autonomous multi-agent coordination may reduce direct human oversight and could amplify harmful actions when assumptions, evidence, or permissions are unreliable or inappropriate.Source: Ethical Considerations · p. 9
  • RCM makes operational responsibility explicit and constrains interventions by roles, authority, and agreed terms, but does not guarantee safe behavior or replace human accountability.Source: Ethical Considerations · p. 9
  • Evaluation requires substantial computation, energy, and emissions; the authors limit costs with fixed evaluation sets, reused traces where appropriate, and three-run reporting.Source: Ethical Considerations · p. 9

A.1 Commitment Lifecycle Statuses

The appendix specifies commitment lifecycle transitions and local message processing, then details the evaluation benchmarks, failure signals, baselines, and implementation settings.

A.1 Commitment Lifecycle Statuses

  • A request records a proposal, while acceptance of the same proposal version establishes a binding commitment; discharge requires the delegator to verify evidence against the agreed outcome.Verified unmet outcomes or breached constraints can trigger violation; authorized termination and expiry have distinct terminal statuses.Source: §A.1 · p. 12, Table 8 · p. 13

A.2 RCP Processing and Transport

  • RCP processing validates identity, version, authority, and required evidence locally; duplicate identifiers are idempotent, and ordering is required only within an affected agreement.Transport bindings may change serialization and delivery while preserving commitment identity, versioning, and correlation.Source: §A.2 · p. 12

A.3 Benchmark Details

  • Claw-Eval scores combine normalized weighted completion and robustness, with safety acting as a multiplicative gate; benchmark scores are averaged across tasks and runs.Robustness is the fraction of error-affected tool types later returning a successful response, or 1 when no error is injected.Source: §A.3 · p. 13
  • AIME25 and HumanEval assess mathematical reasoning and functional code generation, while Claw-Eval tests controlled failures and Claw-Eval-Live tests changing workflows.Claw-Eval uses traces, audit logs, and environment snapshots to assess completion, safety, and robustness; Claw-Eval-Live uses reproducible snapshots and execution evidence.Source: §A.3 · pp. 12–13

A.4 Failure Signals

  • Failure signals are grouped as execution, resource/state, operational, and validation/safety events, then deduplicated within each task and trial.Each method uses its own observed-failure denominator rather than a cross-method union.Source: §A.4 · p. 14

A.5 Baselines

  • The baseline set spans decentralized communication, hand-designed coordination, and automatically optimized coordination methods.Examples include AgentNet, AutoGen, and GPTSwarm; additional attribution and guard-model baselines are used on specified benchmarks.Source: §A.5 · p. 14

A.6 Implementation Details

  • Main experiments use sixteen-agent teams with Qwen3-4B-Instruct-2507 or GPT-5.4; agentic tasks distribute tools and permissions across agents.RCM allows contact with any teammate, with at most two candidates in one round and one discovery round per attempt; configurations are evaluated over three runs.Source: §A.6 · p. 14

Related papers

Loading 2610.21243v1…