acceptodds
Under review as a conference paper at ICLR 2027

TEA: Structurally Representing Arbitrary LLM Agents Collaboration

Abstract

LLM agent collaboration may take many forms, ranging from predefined or dynamic workflows to open-ended interactions, in which agents create and revise tasks and choose when and with whom to work on the fly. Less structured collaboration lets agents organize their own work and adapt to diverse tasks, environments, and changing teammates, but performance also becomes less reliable and harder to understand. Indeed, even if a workflow is predefined, agents may deviate from it during execution, making tracing each agent's contribution difficult. We introduce Task-Environment-Agents (TEA), a flexible structural representation that makes evolving collaboration explicit without prescribing agent reasoning, behavior, or communication order. TEA couples an Agent Interaction Graph (AIG) to both a Task Activity Graph (TAG) and the environment to form a collaboration state, providing a basis for analysis and optimization. We demonstrate TEA’s potential by representing and recovering emergent collaboration structures and analyzing collaboration states. TEA also serves as an active framework through which agents communicate, transform tasks, and act in the environment. We show that TEA enables and improves decentralized agent collaboration on shared tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM-powered Multi-agent Systems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xjGkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces TEA, which links an Agent Interaction Graph, a Task Activity Graph, and a shared environment to record interactions, task evolution, and supporting evidence. It also introduces Infuse, which derives collaboration guidance from execution traces. The experiments include examples of structural representation and recovery, as well as performance evaluations on code refactoring and Silo-Bench.

Critical strengths
  1. Linking task progress to agent interactions has practical value for understanding and debugging open-ended collaboration.

  2. Persistent task identities, version histories, and evidence links provide concrete mechanisms for inspecting past tasks and tracing how work was carried out.

Critical weaknesses
  1. The motivation and distinct technical contribution of TAG are not clearly established. Section 1 lacks clear criteria for "emergent collaboration" and does not adequately distinguish dynamic behavior at runtime from the emergence of new coordination patterns. AgentOrchestra's TEA protocol [1] already addresses unified resource organization, while TalkHier [2] studies structured communication and hierarchical refinement. The paper therefore needs to clarify what TAG uniquely enables. Section 3 defines task versions and associated operations without sufficiently explaining why they are needed. Topology recovery in §4.1 relies primarily on AIG and does not establish a key role for TAG. As the paper moves from representation and recovery to policy improvement, the central insight and the connections between these contributions remain unclear.

  2. The experiments do not directly validate the main contributions. Section 6.1 demonstrates representation, recovery, and traceability, but does not clearly show what TAG adds to identifying or explaining emergent collaboration. In §6.2.1, removing TAG also removes task tools, evidence access, and task observations, so the ablation cannot distinguish the benefits of the graph structure from those of shared task information. The performance gains in §6.2.2 likewise do not establish a link between TAG analysis and effective policy changes. The experiments therefore provide insufficient evidence for the importance of the paper's central design choices.

[1] Zhang, W., et al. (2025). AgentOrchestra: Orchestrating Hierarchical Multi-Agent Intelligence with the Tool-Environment-Agent (TEA) Protocol. arXiv:2506.12508v4.

[2] Wang, Z., Moriyama, S., Wang, W.-Y., Gangopadhyay, B., & Takamatsu, S. (2025). Talk Structurally, Act Hierarchically: A Collaborative Framework for LLM Multi-Agent Systems. arXiv:2502.11098.

Questions for authors
  1. What criteria define "emergent collaboration", and how do the experiments test whether these criteria are met?

  2. Which existing result best demonstrates TAG's distinct value, and how can the benefits of its structure be separated from those of shared task information?

4 comments
AuthorsOpenReview ↗

Response (Part 1 of 2)

We thank the reviewer for recognizing the value of linking task progress to agent interactions and for the suggestions on TAG’s structural contribution. Below, we clarify the setting and present new experiments separating the structural effect from TAG.


W1/Q1: Emergent collaboration

LLM agents can collaborate under different degrees of policy specification. For example, a workflow could prescribe how agents work together, specifying their roles and interaction order. An orchestrator-subagent system may prescribe delegation relationships while letting the orchestrator decide what to delegate. More explicit specifications often require greater domain knowledge and design effort and can limit flexibility across tasks. We use emergent collaboration to describe work organization and coordination arising from agents’ decisions where the initial collaboration policy leaves them unspecified, such as how to divide work, whom to ask for help, when to share intermediate results. Section 3.1 now clarifies these settings (changes in blue).

We view emergent collaboration as a potentially more powerful setting. We believe it is often overlooked partly because it is difficult to study: there is no explicit collaboration procedure to vary and optimize. Our question is: how can we understand collaboration from agents’ actual interactions? TEA represents collaboration under any degree of policy specification. Section 5 explains how agents use TEA’s collaboration interface to organize collaboration themselves.

W1: Prior work (AgentOrchestra, TalkHier)

Both works give agents a collaboration structure to work within.

AgentOrchestra’s Tool-Environment-Agent protocol provides a standard way to use tools, access environments, and invoke agents. Its protocol could complement our Task-Environment-Agents (TEA) by supporting tool calls during AIG activations, where these calls and their results are recorded. AgentOrchestra builds on this protocol with a central planner that divides tasks and delegates work to specialized subagents. Our TEA is a representation framework for understanding collaborative behavior: which agent acted and when (AIG), what it did in the environment (Z), and why, using task goals and progress as context (TAG). Linking these layers makes emergent collaboration analyzable.

TalkHier uses an explicit collaboration structure to improve task outputs. It defines a graph with roles for generating, evaluating, and revising outputs, coordinated by supervisors. Our TEA enables analysis of agents’ collaboration without requiring a structure to be known in advance, while Infuse offers a mechanism for improving collaboration structure towards more effective task completion.

AuthorsOpenReview ↗

Response (Part 2 of 2)

W1: What does TAG add beyond AIG?

We see TEA as a starting point for studying agent collaboration, with further potential to explore. The recovery and relationship analysis in Section 4, together with Infuse in Section 5, illustrate one concrete way to use TEA to understand and improve collaboration.

Topology recovery and candidate rule inference indeed rely primarily on AIG. However, the power of TEA lies in linking AIG, TAG, and the environment to make the collaboration process interpretable and tie it to task success. We have clarified the roles of AIG and TAG in Section 4 (changes in blue). The new experiments below also provide empirical evidence for the value of this structure in collaboration and policy refinement.

W2/Q2: Separating TAG’s structure from shared information

To isolate TAG’s structural contribution, we compare TEA (TAG + communication) with Structural ablation, which hides TAG edges from agents and Infuse. In addition, we compare TEA with Blackboard, which replaces TAG with a shared read/write workspace. The new Section 6.3 and Tables 8-9 report five runs on shop and all 30 Silo-Bench tasks.

TEA with Infuse achieves the strongest overall results. After Infuse, it leads both LoC reduction metrics on shop and all three Silo-Bench metrics.

TAG’s structure makes compact collaboration history useful. In Raw runs, TEA’s TAG record is about one fifth Blackboard’s size at step 100 on shop (5.2k versus 25.2k characters) and less than half its average size on Silo-Bench (2.6k versus 6.3k). TEA and Structural ablation maintain similarly sized records, yet TEA achieves higher performance in all cases, suggesting that explicit task relationships make compact context useful for understanding collaboration.

Long, unfiltered history may make collaboration harder to improve. Blackboard interestingly achieves the highest Raw code reductions, but both reduction metrics decline after Infuse. However, on Silo-Bench, it matches TEA’s Raw success and benefits from Infuse. Blackboard’s record size turns out to be much smaller on Silo-Bench than on shop (6.3k versus 25k characters). Longer, unfiltered histories may make it harder for Infuse to interpret collaborative behavior.

TEA keeps history compact, and its explicit task relationships help agents and Infuse use that history to understand and improve collaboration.


We are happy to provide further clarification and discussion, and we hope you will consider raising your score if your concerns have been addressed.

Thank you for the detailed response and additional experiments. The clarifications and new ablation studies help address my main concerns, particularly regarding TAG's structural contribution beyond shared task information. I recommend that the authors further clarify the connection between TAG structure, emergent collaboration analysis, and policy improvement in the revised paper. I am then willing to raise my score from 2 to 3.

Thank you for recognizing our work and for your positive feedback. We are glad that the additional experiments helped address your main concerns. We have further revised Sections 4.2 and 5 of the manuscript (changes in orange) to make our contributions to analyzing and improving emergent collaboration more explicit. We appreciate your constructive suggestions!

Reviewer_2WZ9OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Task-Environment-Agents (TEA), a unified structural representation for LLM multi-agent collaboration that accommodates both predefined workflows and emergent, asynchronous coordination. TEA combines an Agent Interaction Graph (AIG), which records agent activations and information exchanges, a Task Activity Graph (TAG), which captures the creation, revision, splitting, and merging of tasks, and a shared environment with evidence attachments. The authors establish representational properties for the two graphs and demonstrate how execution traces can be analyzed to recover interaction topologies and infer candidate coordination rules.

The paper further proposes Infuse, an iterative policy-improvement approach that analyzes TEA traces and environmental outcomes to generate collaboration guidance. Validation includes qualitative demonstrations of different collaboration structures, policy-recovery examples, and empirical experiments on six code-refactoring repositories and 30 Silo-Bench tasks. The results suggest that access to TAG can improve coordination and task outcomes, while Infuse provides additional, though relatively modest, performance improvements.

Critical strengths

1. A unified representation of asynchronous interaction and evolving tasks.

TEA's main contribution is the explicit connection between agent interactions, persistent task identities and versions, evidence, and environmental outcomes. Unlike representations focused primarily on communication traces or predefined workflow structures, TEA can record how agents dynamically reorganize shared work and relate those changes to their interactions. This is consequential because it provides a potentially reusable foundation for analyzing collaboration that is not fully specified in advance.

2. Promising empirical evidence for structured task sharing.

The experiments provide evidence that TAG is useful as an active collaboration interface rather than merely a logging mechanism. On the shop refactoring task, TAG combined with communication achieves 43.0% code reduction among passing runs, compared with 19.5% using communication alone. On Silo-Bench, the corresponding success rate increases from 0.60 to 0.77. These results support the practical relevance of explicit task-state sharing and constitute the strongest empirical argument for the paper.

Critical weaknesses

1. The empirical evaluation does not sufficiently establish that the proposed mechanisms outperform simpler alternatives or produce robust improvements.

The four-condition ablation demonstrates the usefulness of TAG access, but it does not compare TEA against simpler shared task lists, blackboard-based coordination, or existing multi-agent frameworks under matched information access and resource budgets. Consequently, it remains unclear how much of the improvement comes from TEA's particular structural design rather than the general availability of a persistent shared workspace.

The additional benefits of Infuse are not yet convincingly established. As shown in Table 6, Infuse improves the average pass rate from 0.63 to 0.70 and the mean code reduction among passing runs from 25.5% to 30.4% across six refactoring repositories. Although code reduction improves consistently across repositories, the changes in pass rates are mixed, including a decrease on payroll. Similarly, Table 7 shows that Infuse improves Silo-Bench success from 0.77 to 0.83 under TAG + communication, but reduces it from 0.77 to 0.73 under TAG only. While these results demonstrate the potential value of Infuse, the varying effects across settings and the absence of further statistical analysis make it difficult to assess the reliability and generality of the improvements. Consequently, stronger evidence is needed to establish that Infuse consistently improves collaboration policies.

Furthermore, Infuse introduces additional LLM-based feedback and policy guidance, making it difficult to separate the benefits of TEA-specific structural analysis from those of generic experience-based reflection. This attribution gap materially weakens the paper's claims about policy optimization.

2. The generality and effectiveness of collaboration-policy recovery remain insufficiently validated.

The theoretical propositions show that AIG and TAG can encode arbitrary finite DAGs under their respective construction assumptions. However, this expressiveness does not by itself establish that TEA can faithfully or efficiently recover arbitrary collaboration policies from finite, noisy, or evolving execution traces.

The proposed recovery procedure primarily aggregates observed message exchanges and uses temporal patterns to infer candidate execution rules. Recovering observed communication edges is relatively straightforward when complete records are available, whereas distinguishing conditional routing, synchronization rules, and genuine policy constraints from incidental behavior is substantially more challenging.

Although the paper provides illustrative recovery examples, it lacks a systematic evaluation of recovery accuracy across diverse policies, observation lengths, concurrency patterns, and policy changes. This matters because reliable structural recovery is one of the central capabilities motivating TEA and the subsequent Infuse optimization procedure.

Questions for authors

1. Clarification — Contribution of TEA-specific structure. Could the authors compare TEA against a simpler shared task registry or blackboard-based collaboration interface, while controlling for available information, agent steps, and LLM inference budgets? This would help establish whether the improvements arise from TEA's structural design rather than persistent information sharing alone.

2. Clarification — Reliability of policy recovery. Can the authors provide quantitative measurements of topology and coordination-rule recovery, including precision/recall for communication edges and accuracy for conditional execution and joins? How sensitive are these results to incomplete traces, the temporal threshold, and changing collaboration policies?

3. Clarification — Robustness and attribution of Infuse improvements. How does Infuse compare with a generic LLM reflection baseline that receives comparable execution feedback but no TEA-specific structural analysis? Are the reported improvements statistically reliable across random seeds, and do the learned lessons transfer to unseen tasks or repositories?

6 comments

We thank the reviewer for recognizing TEA’s value in representing collaboration organized by agents. We address the three questions below.


Q1/Q3: Contribution of TEA-specific structure and Performance of Infuse

To isolate TAG’s structural contribution, we compare TEA (TAG + communication) with Structural ablation, which hides TAG edges from agents and Infuse. In addition, we compare TEA with Blackboard, which replaces TAG with a shared read/write workspace. The new Section 6.3 and Tables 8-9 report five runs on shop and all 30 Silo-Bench tasks.

TEA with Infuse achieves the strongest overall results. After Infuse, it leads both LoC reduction metrics on shop and all three Silo-Bench metrics.

TAG’s structure makes compact collaboration history useful. In Raw runs, TEA’s TAG record is about one fifth Blackboard’s size at step 100 on shop (5.2k versus 25.2k characters) and less than half its average size on Silo-Bench (2.6k versus 6.3k). TEA and Structural ablation maintain similarly sized records, yet TEA achieves higher performance in all cases, suggesting that explicit task relationships make compact context useful for understanding collaboration.

Long, unfiltered history may make collaboration harder to improve. Blackboard interestingly achieves the highest Raw code reductions, but both reduction metrics decline after Infuse. However, on Silo-Bench, it matches TEA’s Raw success and benefits from Infuse. Blackboard’s record size turns out to be much smaller on Silo-Bench than on shop (6.3k versus 25k characters). Longer, unfiltered histories may make it harder for Infuse to interpret collaborative behavior.

TEA keeps history compact and task relationships explicit, helping agents and Infuse understand and improve collaboration. Collaboration discovery and Infuse demonstrate one practical use: refining collaboration when there is no predefined coordination procedure to vary and optimize. We see TEA as a starting point for studying agent-generated collaboration, with further potential to explore. We also appreciate the reviewer’s recommendation on lesson transfer and see it as a promising next step.

Q2: Reliability of policy recovery

We thank the reviewer for the recommendations on recovery. We discuss two conditions that affect the quality of recovery.

First, a communication edge can be recovered only if agents use it and the interaction is recorded. In Appendix D.1, Figure 9, we simulate six agents on a random graph with probabilistic routing. After 10 steps, many edges have not yet been activated. As more interactions occur, the maximum error in the estimated edge probabilities falls from 0.70 at step 10 to 0.13 at step 100.

Second, interpreting a recovered topology and candidate coordination rules requires context beyond the aggregated graph. Figure 6 shows that aggregating interactions across policy phases mixes different collaboration patterns. For example, if a temporal boundary is off by x seconds and the duration of a collaboration structure is y seconds, we expect, on average, an error rate on the order of x/y in assigning interactions to phases. By linking agent activations to tasks and their progress, TEA also offers a way to analyze how interaction patterns vary across tasks and stages of the work. We agree that systematic tests of coordination-rule accuracy, missing records, and timing thresholds would strengthen this analysis and see these as promising directions for further exploration.


We are happy to provide further clarification and discussion, and we hope you will consider raising your score if your concerns have been addressed.

Thank you for addressing my comments and providing further context. I maintain my original score.

Thank you for recognizing our work. Your suggestions on baseline comparisons, policy recovery, and Infuse helped strengthen our discussion and empirical evaluation. We appreciate your time!

Reviewer_5UgeOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper is about providing structure to LLM agents. The paper describes Agent Interaction Graph (AIG) and a Task Activity Graph (TAG) to make Task-Environment Agents (TEA). Then later they introduce "infuse" which does iterative policy improvement. The experimental results show that the TAG, TEA, and infuse systems perform extremely well compared to ablations. However, the authors do not compare against other agentic comms strategies.

Critical strengths
  • I think the representation experiments and Fig 4 and Fig 5 are well done.

  • I think using Git as an agent analogy is wise. I think the whole paper would be better written if it basically used git as a framework and the diverged when Git is not suitable. This way readers can take something they understand (Git) and then apply it and your technical merit would be the parts where Git does not fit.

  • The empirical results looks really good. Although I'm not sure that these experiments are sufficiently comparative. Surely there dozens of other communication frameworks that you could compare against

Critical weaknesses
  • The proposition -> remark structure of the technical parts of the paper are strange and underwhelming. I'm not sure this is the right structure for this paper.

  • Similar to the previous point. Section 3 and 4 are needlessly complex. I think this could be much better and more clearly written in a plain style.

  • Comparative benchmarks are lacking. There are dozens of other comparative benchmarks for agent communication that are not evaluated.

Questions for authors

My principal question of the technical part of the paper is: how does this compare to other formulations. The paper is written so that this TEA, TAG, AIG just kind of exist. They certainly rhyme with other technical constructs, but as the paper is written the technical merit is not satisfying. I dont see a question that needs resolved or a specific contribution that fills some void.

Are there any other agent communication benchmarks that could be compared against in the Silo-Bench and Refactoring challenges?

2 comments
AuthorsOpenReview ↗

Response (Part 1 of 2)

We thank the reviewer for recognizing the representation experiments, the Git analogy, and the empirical results.


Q1: TEA’s contribution

LLM agents can collaborate under different degrees of policy specification. For example, a workflow could prescribe how agents work together, specifying their roles and interaction order. An orchestrator–subagent system may prescribe delegation relationships while letting the orchestrator decide what to delegate. More explicit specifications often require greater domain knowledge and design effort, and their assumptions can limit agent flexibility and generalization across tasks.

When policies leave agents room to decide, for example, how to divide work or when and with whom to share intermediate results, understanding how they collaborate within those guidelines remains difficult but valuable, especially in open-ended settings. We view open-ended collaboration as an emerging and potentially more powerful setting: LLM agents use their reasoning and tool-use capabilities to organize shared work and adapt their coordination during execution. We believe this setting is often overlooked partly because it is difficult to study: there is no explicit collaboration procedure to begin with, and thus no structure to vary and optimize. Our research question is: how can we understand collaboration from agents’ actual interactions and use these insights to suggest more effective collaboration structures for a given task? TEA addresses this question while supporting any degree of policy specification. We added Section 3.1 to clarify these settings.

TEA records how agents interact in AIG and their stated intentions and task progress in TAG, grounding both in environmental changes. This supports interpreting collaboration without assuming a predefined structure (Section 4). Infuse builds on TEA as a general mechanism for refining collaboration policies from this understanding (Section 5). This demonstrates one practical use of TEA; we believe that TEA’s broader potential as a representation framework for agent collaboration warrants further exploration.

W1/W2: Presentation and the Git analogy

Following the reviewer’s suggestion, we expanded Section 3 (changes in blue) with collaboration settings and a refactoring example before the formal definitions. Section 3 is a carefully selected, concise description of TEA’s core components to keep TEA’s components and relationships unambiguous, where full details appear in Appendices A and B.

Git provides intuition for TAG’s versions and transformations. However, there are two aspects of collaboration that Git does not directly capture. (1) Its commit history does not explain how contributors discussed the work or divided responsibilities. (2) Git’s vocabulary is built around repositories and file revisions. Agentic collaboration can also happen elsewhere, for example, when agents divide household chores, change who does what, and carry out actions.

TEA’s value lies in linking interactions, tasks, and environmental changes: we can examine what agents discussed, how they divided or reassigned work, what they did in the environment, and what progress followed. Agents use tools to interact with environments beyond codebases, while TAG keeps task history compact and organized. The propositions establish the expressiveness of this compact interface, whose vocabulary can capture the structures expressed by Git operations.

AuthorsOpenReview ↗

Response (Part 2 of 2)

Q2/W3: Additional communication comparisons

There are indeed frameworks for organizing and optimizing LLM-agent collaboration. LangGraph, AutoGen, and MetaGPT help define collaboration procedures, while GPTSwarm, AFlow, and MaAS optimize explicit designs. However, unlike these prior works, TEA focuses on analyzing collaboration arising at least in part from agents’ own decisions. It does not assume that procedures are known beforehand, but helps identify how collaboration is organized from agents’ actual interactions.

We compare TEA (TAG + communication) with Blackboard, which replaces TAG with a shared read/write workspace, and Structural ablation, which hides TAG edges from agents and Infuse. The newly added Section 6.3 and Tables 8-9 in the updated manuscript (changes in blue) report results from five runs on shop and all 30 Silo-Bench tasks.

TEA with Infuse achieves the strongest overall results. After Infuse, it leads both LoC reduction metrics on shop and all three Silo-Bench metrics.

TAG’s structure makes compact collaboration history useful. In Raw runs, TEA’s TAG record is about one fifth Blackboard’s size at step 100 on shop (5.2k versus 25.2k characters) and less than half its average size on Silo-Bench (2.6k versus 6.3k). TEA and Structural ablation maintain similarly sized records, yet TEA achieves higher performance in all cases, suggesting that explicit task relationships make compact context useful for understanding collaboration.

Long, unfiltered history may make collaboration harder to improve. Blackboard interestingly achieves the highest Raw code reductions, but both reduction metrics decline after Infuse. However, on Silo-Bench, it matches TEA’s Raw success and benefits from Infuse. Blackboard’s record size turns out to be much smaller on Silo-Bench than on shop (6.3k versus 25k characters). Longer, unfiltered histories may make it harder for Infuse to interpret collaborative behavior. TEA keeps history compact, and its explicit task relationships help agents and Infuse use that history to understand and improve collaboration.

We also welcome the reviewer’s suggestions of other frameworks to compare against.


We are happy to provide further clarification and discussion as needed, and we hope you will consider raising your score if your concerns have been addressed.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Overall Summary

Dear Reviewers and AC,

We thank the reviewers for their thoughtful feedback. Below, we summarize our contributions, recognized strengths, and our responses to the major concerns. We also updated the manuscript, with changes highlighted in blue.

Highlights of TEA.
LLM agents can collaborate under different degrees of policy specification. For example, an agent may carry out its role as a node in a predefined workflow that determines how it collaborates with others. However, specifying these roles and interactions often requires domain knowledge and design effort, and these assumptions can limit flexibility across tasks. Alternatively, agents may decide how to collaborate as they complete a task, including what to do and what to communicate. When some or all of these decisions are left to agents, collaboration becomes difficult to understand and improve. TEA aims to understand collaboration from observations of agents’ decisions and behaviors and use this understanding to improve collaboration.

TEA makes the following novel contributions.

  1. We introduce Task-Environment-Agents (TEA), a framework for interpreting collaboration through a unified structural representation of three layers: agent interactions through information exchange (AIG); task formation, responsibility assignment, and progress (TAG); and environmental changes.
  2. We show TEA’s potential through concrete examples of discovering interaction topologies, inferring candidate coordination rules, and analyzing collaboration through the structural relationships it records. We show how TEA identifies collaboration structures from observed agent behavior alone, without prior knowledge of them.
  3. We implement TEA as a collaboration interface, exposing its components as tools to communicate, update tasks, and act in the environment. We introduce Infuse, which uses TEA’s representation to guide agent behavior for better collaboration on specific tasks.

Recognized Strengths.

  1. TEA’s unified representation across three layers. Reviewers 2WZ9 and xjGk recognized TEA’s value in linking agent interactions, evolving tasks, and environmental outcomes to understand collaboration that is not fully specified in advance.
  2. Structural Value of TEA. Reviewer 5Uge highlighted the experiments showing that TEA could distinguish different collaboration patterns and the comparison with Git. Reviewer xjGk recognized persistent task identities and version histories as mechanisms for understanding how work was carried out.
  3. Results. Reviewer 2WZ9 highlighted TAG’s role as an active collaboration interface and its improvements over communication alone in both environments. Reviewer 5Uge also recognized the promising empirical results.

Addressing the Major Concerns.

  1. Understanding the structural benefits of TEA. Reviewers xjGk and 2WZ9 suggested further isolating TEA’s structural benefits. We added Section 6.3 with an ablation that hides TAG edge information from agents during execution and Infuse during policy refinement. The resulting histories are similar in size, but TEA outperforms the ablation in all cases. These results suggest TEA’s structural information is key to keeping collaboration history useful and actionable.
  2. Blackboard baseline. To further understand TAG’s benefits and compare it with other communication methods (as suggested by Reviewer 5Uge), we followed Reviewer 2WZ9’s suggestion and added the Blackboard baseline. It replaces TAG with a shared read/write workspace. TEA with Infuse achieves the best overall results on both tasks. One caveat is that Blackboard achieves higher code reduction on the code-refactoring task before Infuse, but its performance drops after Infuse. We found that Blackboard’s history grows almost linearly over time, whereas TAG’s history is about one fifth as large at step 100 on the code-refactoring task and less than half as large on Silo-Bench. These results suggest that long, unfiltered histories can make it harder to extract actionable insights for improving collaboration.
  3. Comparison with Git. Following Reviewer 5Uge’s suggestion, we clarified how Git provides intuition for TAG’s compact but expressive vocabulary for representing work history, and that TEA applies beyond codebases to record what agents discussed, how they divided work, and changes in these decisions over time.
  4. Policy recovery. We further discussed how TEA recovers interaction topologies (Reviewer 2WZ9). With one underlying interaction topology, more observations can improve recovery accuracy. When collaboration spans multiple phases, each with its own topology, recovery quality depends on whether the observation window overlaps multiple phases and, if so, on the overlap duration and the frequency of interactions from each phase.

We appreciate the reviewers’ feedback, which helped make TEA easier to follow and strengthen the evidence for its structural benefits.

Sincerely, Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.