acceptodds
Under review as a conference paper at ICLR 2027

NanoMA: How Does Agent ReAct Enables Flexible Multi-Agent Topologies?

Abstract

Long-horizon tasks continually reveal new dependencies and evidence, requiring multi-agent collaboration to adapt during execution. Central planning concentrates coordination in one controller, while recursive delegation distributes task decomposition but offers limited support for revising existing cross-branch collaboration. We introduce NanoMA, a multi-agent ReAct (Reasoning + Acting) framework that treats collaboration structure as an executable and dynamically revisable component of problem solving. Within each agent's ReAct loop, execution actions are combined with four topology actions—spawn, kill, query, and send—for creating, terminating, observing, and contacting agents. Execution traces show agents combining these actions in response to task feedback to form diverse collaboration structures, rather than following predefined topology templates. On RepoZero-Py2JS Hard with GPT-5.6 Luna High, NanoMA achieves a 66.91% all-pass rate, exceeding the best-performing baseline in our comparison by 13.67 percentage points. On EdgeBench, it obtains the highest overall score among the evaluated frameworks with both GPT-5.6 Luna and Qwen3.7-Plus. Communication ablations show higher task accuracy with cross-branch query access and better performance with targeted messages than with broadcasting. Code is available anonymously at https://anonymous.4open.science/r/NanoMA-private-2CC6/.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
LLM agents, multi-agent systems, dynamic orchestration, decentralized planning, long-horizon tasks, agent communication, dynamic collaboration topology

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_784hOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NanoMA lets agents create, stop, query, and message other agents during task execution, so the team structure can change as work progresses.

Critical strengths

NanoMA reaches high all-pass on RepoZero, much above the baseline.

Critical weaknesses
  1. It changes queries and message recipients, but does not isolate creating or stopping agents. Better communication could help even with a fixed team. The ablations does not establish the benefit of changing team membership. Compare with fixed teams using the same messaging, queries, and runtime.

  2. Table 7 reports 3.53M tokens for NanoMA versus 0.43M for CrewAI. Extra attempts and checks could explain part of the gain. The current advantage are difficult to attribute to coordination alone. Compare success at equal token budgets and report cost per successful task.

  3. It adds validation, rollback, a durable result ledger, and publication checks alongside flexible teams. These features could improve results without changing team membership. This bundled design make the source of NanoMA’s gain unclear. A fixed-team control should retain the same ledger, workspace rules, and validation.

  4. It counts send calls, but one broadcast can expose many agents to the same text. Similar call rates therefore do not mean equal received context. The current comparison does not separate recipient choice from extra distracting messages. This distinction need received-token counts and comparisons with similar message volume and team size.

  5. It allows later overlapping edits to overwrite earlier changes. Rollback occurs when registered checks detect failure, so a lost requirement could survive if those checks miss it. The paper does not measure this failure rate. The current validation process do not establish that parallel work is preserved. Test overlapping patches with known requirements and report overwritten successes and recovery.

Questions for authors

Please refer to the weaknesses.

21 comments
AuthorsOpenReview ↗

W1-a: Spawn and Inter-Agent Communication

Thank you for your suggestions on separating orchestration and communication effects.

W1-a: Spawn and Inter-Agent Communication

To address your question about separating agent creation from communication, completely disabling spawn in NanoMA’s single-root initialization leaves only one agent and also removes inter-agent communication. The no-spawn ablation therefore changes both team formation and the opportunity to communicate; removing this tool alone cannot independently measure their contributions. We consequently use prescribed-topology controls that retain multiple agents and default communication capabilities to examine the effect of team organization.

AuthorsOpenReview ↗

W1-b: Prescribed Topologies with Default Communication

W1-b: Prescribed Topologies with Default Communication

To address your request for team comparisons with the same messaging, queries, and runtime, we write skills that constrain spawn-tool calls so agents can only be created according to the prescribed star, transitive-tournament, or strict-binary-tree structure, while retaining default query/send permissions, the model, runtime, and resource limits. These controls preserve inter-agent communication while constraining team organization. Their All-pass rates are 51.80%, 55.40%, and 51.08%, and Micro scores are 79.43%, 80.80%, and 78.32%, respectively, versus 66.91% All-pass and 93.32% Micro for autonomous NanoMA. Actual team sizes and token use are reported below.

Prescribed orchestration with default communication

OrchestrationAll-passMicroTotal tokensMean agents
Default NanoMA · autonomous93/139 (66.91%)93.32%493.69M6.82
Prescribed star · centralized72/139 (51.80%)79.43%197.04M3.51
Prescribed transitive tournament77/139 (55.40%)80.80%267.43M4.17
Strict binary tree · local71/139 (51.08%)78.32%218.01M5.01

Each prescribed topology is implemented through skills that constrain spawn-tool calls so that agent creation follows the specified structure. All configurations retain default query/send permissions and use the same backbone, runtime, and resource limits. All-pass uses 139 tasks, counting missing runs as failures. Micro uses the same aggregation as the default result. Topology definitions are provided in the main text and appendix. Actual token use and team sizes are reported, not assumed equal.

AuthorsOpenReview ↗

W1-c: Spawn Ablation

W1-c: Spawn Ablation

To address your question about agent creation, we compare autonomous spawning, forced spawning at every task_create, and single-agent execution with spawn disabled on shorter Terminal-Bench 2 tasks and longer EdgeBench tasks. Terminal-Bench 2 tests whether selective delegation also benefits shorter tasks that may require only simple collaboration. Within each benchmark, the model, runtime, and budget settings are identical. Autonomous spawning scores highest on both benchmarks, while forced expansion performs worse than single-agent execution, supporting the value of deciding whether and when to delegate.

Agent-creation policy ablation

ConfigurationSpawn policyTerminal-Bench 2 · Luna xhighEdgeBench
Default NanoMAAutonomous decisions57.30% (51/89)22.10
Forced spawningSpawn at every task_create41.57% (37/89)16.29
Spawn disabledSingle agent53.93% (48/89)18.21

*Within each benchmark, all three configurations use the same backbone, runtime, and budget settings. EdgeBench values are scores, not percentages.

Reviewer_71DFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NanoMA, a multi-agent framework that enables dynamic collaboration through four topology actions: spawn, kill, query, and send. By integrating these actions into each agent’s ReAct loop, NanoMA allows agents to autonomously adjust their collaboration structures. Experiments on RepoZero-Py2JS Hard and EdgeBench demonstrate promising improvements over existing frameworks.

Critical strengths
  1. The paper addresses an important problem in multi-agent collaboration by enabling agents to dynamically adjust their collaboration structures during execution.
  2. The experimental results are promising, with substantial improvements over several baselines on RepoZero-Py2JS Hard and competitive performance on EdgeBench.
Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily integrates existing agent management and communication operations (spawn, kill, query, and send) into the ReAct loop. The paper does not sufficiently clarify its fundamental differences from existing dynamic multi-agent frameworks.
  2. Insufficient ablation studies. The experiments only investigate query and send, without isolating the contributions of spawn and kill. Consequently, it remains unclear whether the improvements primarily result from dynamic topology adaptation or enhanced inter-agent communication.
Questions for authors

Can the authors provide additional ablation experiments to isolate the contributions of spawn and kill, and clarify how much of the performance improvement can be attributed to dynamic topology adaptation rather than inter-agent communication?

10 comments
AuthorsOpenReview ↗

W1-a: Task-Driven Topology Adaptation

Thank you for your interest in the novelty of our paper.

W1-a: Task-Driven Topology Adaptation

To address your question about methodological novelty, we highlight our core contribution: enabling the model to choose, form, and revise diverse collaboration topologies as task needs emerge. As you may have noted, Figure 1 compares centralized planning through a controller, recursive delegation through parent–child hierarchies, and NanoMA’s node-autonomous orchestration. Each NanoMA node combines execution and orchestration decisions in its ReAct loop; creation links record provenance, while collaboration can extend across branches. This makes topology adaptation part of task execution.

Figure 4 records 12 structural labels for NanoMA, 5 for CC-Workflow, and 2 for Omnigent; these describe different cohorts and include whole-graph patterns and local motifs, not capability limits. Skills also allow prescribed orchestration: in the fixed-topology controls, we write skills that constrain spawn-tool calls so agent creation follows the specified structure. With default communication and runtime retained, autonomous NanoMA achieves 66.91% All-pass / 93.32% Micro, versus 51.80% / 79.43%, 55.40% / 80.80%, and 51.08% / 78.32% for star, transitive-tournament, and strict-binary-tree policies.

AuthorsOpenReview ↗

W1-b: Adaptation in Concrete Tasks

W1-b: Adaptation in Concrete Tasks

To address your question about differences from existing dynamic multi-agent frameworks, we added concrete traces of nodes organizing work as new needs emerge. In jsonschema/test9, bravo identified a need for independent Draft7 review and created echo, selectively deepening the validator branch. Echo sent semantic recommendations to both bravo and alpha; they incorporated them into implementation and integration, then validated the corresponding behavior. The executing node thus organized specialist support without requiring the root to predefine that role.

In schedule/test13, new CLI requirements prompted delta to create echo for implementation and foxtrot for independent verification. Foxtrot sent compatibility findings directly to alpha, which adopted them and verified the changes. These cases show local reorganization and downstream use. The star control also reaches 100% on schedule/test13; the observed score differences are not attributed entirely to individual topology actions.

Case-study outcomes on the same tasks

TaskDefault NanoMAPrescribed starPrescribed transitive tournament
jsonschema/test970/70 (100%)51/70 (72.86%)0/70 (0%)
schedule/test13153/153 (100%)153/153 (100%)0/153 (0%)

*Entries are test-case pass counts and rates within each task. These controls use skills to constrain spawn calls to the prescribed topology and specify communication targets; they are distinct from the controls retaining default communication.

AuthorsOpenReview ↗

W1 Conclusion

W1 Conclusion

For your question about novelty, the evidence identifies node-level, task-driven topology adaptation as the methodological contribution: each agent can organize specialist work and revise collaboration as needs emerge, while skills can specify an orchestration policy. The 12 observed labels describe structural variety, and the cases show how local decisions change implementation and validation. The stronger performance with default communication and runtime retained connects this capability to task outcomes, supporting its role as a key contributor to NanoMA’s performance.

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

This paper introduces NanoMA, a multi-agent ReAct framework that supports dynamic collaboration structures. The method distinguishes agent-creation links from collaboration links and uses four actions—spawn, kill, query, and send—to support recursive delegation, cross-branch information access, targeted communication, and agent termination. The paper evaluates task performance and resource consumption on RepoZero-Py2JS Hard and EdgeBench, and examines different collaboration patterns through communication ablations and execution-trace analysis.

Critical strengths

S1. The paper addresses the practically meaningful problem of dynamic multi-agent collaboration in long-horizon tasks.

S2. Separating creation links from collaboration links is intuitively motivated and supports cross-branch information access and targeted communication.

S3. The experiments cover two benchmarks and multiple model backbones, demonstrating the method’s effectiveness on the evaluated tasks.

Critical weaknesses

W1. The current ablations compare query and send communication policies but do not compare local and centralized topology control under the same runtime, communication capabilities, and resource limits. The experiments also do not separately examine the effects of dynamically spawning and terminating agents.

W2. The topology diversity analysis uses different task cohorts and combines relations accumulated over entire executions with local graph patterns. These results show that the system produces diverse structures, but do not directly explain how these structures help it respond to newly emerging task dependencies.

W3. Table 7 reports that NanoMA uses 490.69M tokens with GPT-5.6 Luna High, compared with 60.15M for CrewAI and 266.86M for Omnigent. Although NanoMA outperforms CC-Workflow using fewer tokens, its advantage over other baselines under matched token or monetary budgets remains unverified.

W4. The OpenHands-bash all-pass rates in Table 1 are 44.83% and 43.52%. If these results come from a single evaluation of 139 tasks, neither can be obtained by dividing an integer number of successful tasks by 139 and rounding the resulting percentage to two decimal places.

Questions for authors

Please refer to the weaknesses above.

9 comments
AuthorsOpenReview ↗

W1-a: Centralized and Local Control

Thank you for your suggestions on orchestration comparisons and tool ablations.

W1-a: Centralized and Local Control

To address your question about centralized versus local topology control, we implemented three prescribed policies within NanoMA by writing skills that constrain spawn-tool calls so agent creation follows the specified topology. The star represents centralized orchestration around the root; the strict binary tree represents local orchestration within a prescribed branching structure. The transitive tournament adds a prescribed ordered-coordination structure, testing the comparison beyond these two organizations. All three are compared with default NanoMA, which forms and adjusts its topology according to task needs. The model, runtime, default query/send capabilities, and resource limits are identical across configurations; topology definitions follow the paper and appendix.

Default NanoMA achieves 66.91% All-pass and 93.32% Micro. Against the centralized star, its gains are 15.11 and 13.89 percentage points, respectively; against the local binary tree, 15.83 and 15.00 points; and against the transitive tournament, 11.51 and 12.52 points. Therefore, under the matched runtime, communication capabilities, and resource limits in your question, NanoMA outperforms the evaluated centralized-star and local-binary policies, as well as the additional ordered control.

Prescribed orchestration with default communication

OrchestrationAll-passMicroTotal tokensMean agents
Default NanoMA · autonomous66.91%93.32%493.69M6.82
Prescribed star · centralized51.80%79.43%197.04M3.51
Prescribed transitive tournament55.40%80.80%267.43M4.17
Strict binary tree · local51.08%78.32%218.01M5.01

Each prescribed topology is implemented through skills that constrain spawn-tool calls so that agent creation follows the specified structure. All configurations retain default query/send permissions and use the same backbone, runtime, and resource limits. All-pass uses 139 tasks, counting missing runs as failures. Micro uses the same aggregation as the default result. Topology definitions are provided in the main text and appendix. Actual token use and team sizes are reported, not assumed equal.

AuthorsOpenReview ↗

W1-b: Spawn and Kill Ablations

W1-b: Spawn and Kill Ablations

To address your question about agent creation and termination, we compared autonomous spawning, forced spawning at every task_create, and single-agent execution with spawn disabled. Terminal-Bench 2 tests whether selective delegation remains useful on shorter tasks that may need only simple collaboration; EdgeBench provides the longer-task comparison. Within each benchmark, model, runtime, and budget settings are identical. On Terminal-Bench 2, the three configurations score 57.30%, 41.57%, and 53.93%; on EdgeBench, 22.10, 16.29, and 18.21. Autonomous spawning leads on both, while forced expansion scores below the single-agent configuration.

Agent-creation policy ablation

ConfigurationSpawn policyTerminal-Bench 2 · Luna xhighEdgeBench
Default NanoMAAutonomous decisions57.30%22.10
Forced spawningSpawn at every task_create41.57%16.29
Spawn disabledSingle agent53.93%18.21

Within each benchmark, all three configurations use the same backbone, runtime, and budget settings. EdgeBench values are scores, not percentages.

For kill, we focused on 11 tasks involving kill calls because its infrequent use could dilute differences across the full benchmark. The enhanced variant adds a periodic check: every four agent-loop cycles, agents are prompted to inspect other nodes and decide whether to invoke kill. The tasks were selected from this enhanced run; the disabled variant removes kill from the default configuration. Removing kill from the default reduces the mean score from 26.12 to 25.27 on this subset.

Kill analysis on the selected 11 tasks

MetricDefaultEnhanced killDefault without kill
Mean score26.1221.8725.27
Total kill calls12450

The 11 tasks were selected based on kill calls under an enhanced inspection policy. The tool-removal ablation compares the default configuration with the default minus kill; the enhanced policy is shown for reference.

AuthorsOpenReview ↗

W1 Conclusion

W1 Conclusion

Therefore, under the same runtime, communication capabilities, and resource limits you requested, NanoMA retains its performance advantage: 66.91% All-pass and 93.32% Micro, versus 51.08%–55.40% and 78.32%–80.80% across the centralized-star, local-binary-tree, and additional ordered transitive-tournament policies. These comparisons support autonomous dynamic topology as a key source of NanoMA’s core performance advantage. Execution nodes can organize specialist support and revise collaboration as task requirements emerge, without being confined to the evaluated structural templates. The separate spawn comparisons further address the value of dynamic creation. On shorter Terminal-Bench 2 tasks, autonomous delegation improves over single-agent execution by 3.37 percentage points; on longer EdgeBench tasks, the gain is 3.89 points. Forced spawning underperforms single-agent execution on both, demonstrating that simply expanding the team does not produce the same benefit. Together, the controls under these common conditions and separate tool ablations support letting the model decide whether, when, and where to delegate and how to organize collaboration, across both shorter and longer tasks. We will emphasize this adaptive orchestration as the core advantage, while describing kill as an auxiliary optimization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.