acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Multi-Agent Collaboration: When More Is Less

Abstract

The rapid advancement of large language models and single-agent harnesses has reshaped the landscape of autonomous systems, raising a critical question of when multi-agent collaboration offers genuine value. As individual agent capabilities continue to scale, multi-agent collaboration faces diminishing returns while incurring growing context overhead. Through systematic analysis, we delineate the capability boundaries of multi-agent collaboration relative to single-agent alternatives, showing that it confers systematic benefits specifically in long-horizon tasks with sparse dependencies, while single-agent harnesses remain superior in tightly coupled, sequential workflows. Building on these insights, we propose SAIGE, a lightweight multi-agent collaboration mechanism based on Semantic-Aware Incremental Graph Evolution. SAIGE models collaboration as a dynamically evolving graph, where nodes are agent instances spawned on demand and edges encode semantic dependencies established through content-based information retrieval. Experiments on long-horizon, complex task benchmarks show that SAIGE achieves a favorable trade-off between context efficiency and task performance, and that scaling the agent pool or deepening the recursion level does not consistently improve outcomes. Our findings suggest that multi-agent superiority is bounded by task structure rather than universal, and that more agents do not necessarily make a system more intelligent.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Graph Theory, Agent Orchestration, Scalability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SyGrOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when multi-agent collaboration is beneficial relative to a strong single-agent harness. It argues that the benefit of collaboration depends on task dependency structure: multi-agent systems can help on long-horizon tasks with sparse dependencies, while decomposition may hurt tightly coupled sequential workflows because of coordination overhead and dependency errors. The paper formalizes execution trajectories as dependency DAGs, identifies bridge edges as candidate subtask boundaries, and proves a context-cost lemma under oracle decomposition assumptions. It then proposes SAIGE, which incrementally spawns agents when semantic dependency boundaries are detected and retrieves inter-agent information on demand. Experiments across four long-horizon benchmarks show that SAIGE performs particularly well on sparsely dependent tasks, while avoiding much of the degradation of generic multi-agent baselines on tightly coupled tasks.

Critical strengths

The paper focuses on an important question: when does multi-agent collaboration provide meaningful benefits over a strong single-agent harness? Its dependency-aware perspective is also useful: modeling execution trajectories as dependency graphs and using incremental delegation with on-demand message fetching offers a principled way to relate task structure to context isolation and coordination costs.

Critical weaknesses

1. Overstated theoretical contributions. The paper’s characterization of its results as “core theoretical contributions” overstates what the analysis establishes. It provides a graph-based formalization and a context-cost lemma under oracle decomposition assumptions, but these do not establish when collaboration benefits outweigh coordination costs or substantiate the claimed capability boundaries of multi-agent systems. The analysis also does not justify the claimed advantage of incremental orchestration: Eq. (6) follows from event inclusion rather than demonstrating improved prediction accuracy from additional execution evidence. The theoretical claims should therefore be narrowed to match the results actually established.

2. Insufficient engagement with prior work on when and why MAS helps. The related work focuses primarily on recent LLM-based architectures and topologies, without adequately positioning the contribution against established research on task decomposition, coordination, and the benefits and costs of multi-agent problem solving. Consequently, it remains unclear which insights are new to LLM agents and which restate existing principles.

Questions for authors
  1. In Table 1, SAIGE sometimes uses fewer input tokens than the single-agent baseline, despite introducing additional agents, delegation steps, and inter-agent communication. This seems a little surprising, so maybe clarifying how token usage is measured and aggregated across agents is necessary.

  2. All methods are evaluated with the same foundation model, DeepSeek-V4-Pro. Are the observed advantages of SAIGE robust across different foundation models, or could they be specific to the model’s instruction-following, context-management, or tool-use behavior? Cross-model experiments would be important for theory establishment.

  3. The paper uses Nmax = 4 as the default configuration, but does not explain why this value was selected. The ablation study only compares Nmax = 4, 5, 6, and therefore does not establish whether four agents are actually optimal or simply the smallest tested budget.

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

The paper studies when multi-agent decomposition helps relative to a single agent. It models an execution trajectory as a dependency DAG and shows that splitting a known trajectory along bridge edges can reduce a specified context cost. Based on this view, SAIGE spawns subagents incrementally and fetches their information when needed. All methods use the same Codex CLI harness and foundation model. Across four long-horizon benchmarks, SAIGE stays close to the single agent on Terminal Bench and NL2Repo and improves the reported scores on Deep Research Bench II and AgentIF-OneDay. Ablations examine agent limits, recursion depth, and message fetching. The paper argues that collaboration benefits depend on task structure and that adding agents does not consistently improve performance.

Critical strengths
  1. The evaluation uses demanding agent tasks and a consistent harness and foundation model across methods. Table 1 shows a substantial AgentIF-OneDay gain from 53.57 to 71.80 while input tokens decrease slightly. It also honestly shows that SAIGE does not beat the single agent on Terminal Bench. Reporting both input and output tokens makes the trade-off easier to assess.

  2. The ablations go beyond a single default comparison. Tables 2 and 3 test agent count, recursion depth, and message fetching, and Appendix B.3.2 also relaxes the limits for TDAG and DynTaskMAS. Figure 5 reports actual agent use, with SAIGE mostly retaining single-agent execution on the two coding benchmarks. These results support the practical value of selective delegation.

Critical weaknesses
  1. My main concern is that Equation 6 does not justify the reliability advantage claimed immediately after it. Its identity and inequality follow from conditioning on the earlier set being correct, given the nested predicted edge sets. I do not see a comparison between an incremental predictor and an upfront predictor, or an argument establishing that waiting for more evidence improves either predictor. Section 3.1.3 correctly limits the cost lemma to an idealized known trajectory. My concern is that the stronger performance boundary asserted in the abstract and introduction does not follow from that limited result. I would like the authors to preserve the stated scope and either give an additional argument for incremental reliability or present it as an empirical hypothesis. I consider this a fixable issue in the theoretical claims.

  2. I do not think the bridge-edge evaluation directly validates the formal dependency definition. Appendix B.3.3 labels successful single-agent trajectories using DeepSeek-V4-Pro judgments about whether removing a step would change an action or observation, whereas Equation 1 defines an edge through a thresholded drop in action value. DeepSeek-V4-Pro then matches the multi-agent trajectories to these reference components. A valid alternative workflow can have different boundaries, and the paper does not explain the treatment of unmatched tasks or components. I find Figure 6 useful as agreement evidence, but I do not think the validation supports calling it recovery of the true dependency boundaries and using it to explain performance. I see this as a fixable gap in the evaluation and its interpretation.

Questions for authors
  1. I would like the authors to explain what assumption or comparison in Equation 6 connects the conditional correctness identity to an advantage over upfront prediction. I would also like them to distinguish the claims formally established by the context-cost lemma from hypotheses supported only by the experiments.

  2. I would like the authors to state how many tasks in Figure 6 have a successful single-agent reference and explain how alternative workflows and unmatched components are scored. I would also like to know whether any existing validation connects the semantic dependency labels to the action-value criterion of Equation 1.

Reviewer_VQ6TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates when and why multi-agent collaboration outperforms a well-equipped single-agent harness. The authors analyze the capability boundaries of multi-agent collaboration through graph theory and propose SAIGE, which represents collaboration as a dynamically evolving graph. Nodes correspond to agent instances spawned on demand, while edges encode semantic dependencies established through content-based information retrieval. The theoretical analysis suggests that multi-agent collaboration is beneficial for long-horizon tasks with sparse dependencies, whereas single-agent execution is preferable for tightly coupled, sequential tasks. Experiments across four benchmarks show promising results, with SAIGE outperforming the evaluated baselines on three benchmarks, particularly AgentIF-OneDay. Ablation studies further indicate that increasing the agent budget or recursion depth does not consistently improve performance and often increases token consumption.

Critical strengths
  1. Relevant research question. As single-agent systems become increasingly capable, understanding when multi-agent collaboration provides additional benefits is an important question. Examining this question through task dependency structures offers an interesting perspective.
  2. Promising empirical results and informative ablations. SAIGE outperforms the evaluated baselines on three of the four benchmarks. The ablation studies also provide useful observations about the effects of agent budget and recursion depth, showing that additional agents or deeper recursion do not necessarily lead to better performance.
Critical weaknesses
  1. Insufficient description of the execution workflow. The paper does not clearly explain how SAIGE operates during execution and lacks even a pseudocode-level description of its core procedure. In particular, it remains unclear when and how bridge-edge predictions are made and incorporated into the evolving graph. These details are essential for understanding and evaluating the proposed method.
  2. Insufficient support for the reliability claim. Equation (6) does not sufficiently substantiate the claim that deferring edge prediction until execution evidence accumulates is more reliable than committing to a topology upfront. The connection between the equation and the claimed reliability advantage requires further explanation or direct empirical evidence.
  3. Limited evaluation across foundation models. All experiments use DeepSeek-V4-Pro as the foundation model. Additional experiments with other foundation models would help establish whether the observed benefits generalize beyond this particular model.
Questions for authors
  1. Could the authors provide pseudocode for the SAIGE execution workflow used in the experiments, including when and how bridge edges are predicted and incorporated into the graph?
  2. Could the authors directly compare execution-informed dynamic edge prediction with a topology determined and fixed before execution, using the same foundation model, comparable token budgets, and otherwise consistent execution settings? Such a comparison would help establish whether deferred edge prediction improves dependency identification reliability and whether this translates into better task performance.
2 comments
AuthorsOpenReview ↗

Reminder For Discussion

Dear Reviewer VQ6T, We hope this message finds you well. We would greatly appreciate it if you could take a moment to review our response and consider providing your feedback before the discussion period ends on Nov.18th. Thank you very much for your time and valuable feedback. We look forward to hearing from you.

AuthorsOpenReview ↗

Reminder For Discussion (。◕‿◕。)

Dear Reviewer VQ6T

We hope this message finds you well. We would greatly appreciate it if you could take a moment to review our response and consider providing your feedback before the discussion period ends on Nov.18th. Thank you very much for your time and valuable feedback. We look forward to hearing from you.

Sincerely,

Authors = =

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their responses to the questions and concerns raised in the reviews. Please review their responses and provide any further comments or clarifications, if applicable.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.