Multiagent Systems · arXiv:2610.19342 v2 · submitted 14 Oct 2026 · 49 references
Why Do Agent Teams Fail Across Boundaries? A Failure Taxonomy for Agent-to-Agent Systems with Private States
The paper says
“Agents are increasingly interacting with other agents that act for different principals (people or organizations on whose behalf they work), such as a buyer's personal agent negotiating with an enterprise's seller agent.”
more of the abstract
In such agent-to-agent (A2A) scenarios, each agent has its own environment and holds information and authority that the other cannot see. Moreover, the interests of each agent may be aligned (cooperative), partly aligned (mixed-motive), or opposed (competitive). Existing multi-agent system failure studies assume one orchestrator, shared workspace, or single principal. We ask how A2A systems fail when these assumptions are removed. We study more than 2,500 recorded runs of A2A systems built from different agent harnesses and models, spanning 488 scenarios across eight families (legal, CRM, security, commerce, etc.). From these traces, we derive A2-Failures, a taxonomy of 20 A2A failure modes in four categories: (i) discovery and delegation, (ii) grounding and record consistency, (iii) verification, and (iv) conduct and fidelity. Three findings hold across every harness, model, and scenario studied: (1) discovery and interaction are driven by the phrasing of the task rather than the availability of peer agents; (2) no agent verifies a peer's factual claim on its own initiative, so an A2A system inherits the reliability of its least reliable participant unless the protocol or mechanism performs the check; and (3) the aggregate outcome score credits the agent completing the task while staying blind to the coordination, verification, and conduct failures beneath it, so A2A systems must be evaluated per agent, against each agent's own principal. We release the run logs behind the taxonomy at https://huggingface.co/datasets/a2a-salesforce/a2-failures.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-structured empirical taxonomy of 20 failure modes in agent-to-agent systems with private states and separate principals, with three provocative cross-cutting findings—but the zero-verification claim should be read as an observation under current models and prompts, not a universal property. This line is wrong
1 major objection · 8 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 49 references say.
The paper says, at Abstract and §1, §3 (Findings 2), §6 Limitations
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.60):
“Harness and model effects are therefore not fully separable, and our claims concern which failures occur rather than their rates across configurations.”
Pith says: Objection 1 of 1 · major
The claim that 'no agent verifies a peer's factual claim on its own initiative' is presented as holding 'across every harness, model, and scenario studied.' However, §6 acknowledges that harness and model effects are not fully separable because each harness has a default system prompt that is closely integrated with its model.
more of objection 1
Since all three tested harnesses are closed-source and their system prompts likely do not instruct agents to independently verify peer claims, the observed absence of verification could be an artifact of prompt design rather than a universal property of A2A systems with private states. The paper's own examples of protocol-enforced verification (Appendix D, sha256 checkpoint) show that agents do verify when the mechanism requires it. To avoid overgeneralization, the authors should explicitly discuss this specific confound in §6 and qualify the claim in the abstract and findings to reflect that the result holds under the tested prompting regimes. The rest of the taxonomy and other findings are not affected by this concern.
8 minor objections
The paper says, at §2.2
Pith says: minor objection 1 of 8
The description of the four MCP tools is clear, but the text would benefit from an explicit statement that the protocol does not specify any verification requirement for messages, to explain why verification failures can occur.
The paper says, at §6
Pith says: minor objection 2 of 8
The limitations section states that 'harness and model effects are therefore not fully separable,' but it does not discuss how this specifically limits the generalization of the zero-verification finding. Please add a sentence making this connection explicit.
show 6 more
The paper says, at §3, Insight 3
Pith says: minor objection 3 of 8
The claim 'We found no instances of agents re-fetching a document or re-querying a source to test a peer's factual claim' is a strong negative. The search procedure for such instances should be described more explicitly (e.g., what automatic tags were used, how many traces were manually inspected).
The paper says, at Table 2 and Appendix B
Pith says: minor objection 4 of 8
Several failure modes (e.g., D2, G2) list a 'protocol' layer as suggested intervention, but the paper's own evidence shows that protocol-level changes (e.g., structured capability fields) could help. Clarify whether these interventions are speculative or supported by the trace data.
The paper says, at Appendix D
Pith says: minor objection 5 of 8
The verbatim transcripts are valuable. However, some transcript labels (e.g., 'G1 The disclaimer that makes the gap visible') do not correspond to the failure mode definition in Table 2; G1 is 'Misdirected request,' but the example shows a peer correctly disclaiming missing files. Either relabel or explain how this is a failure.
The paper says, at General
Pith says: minor objection 6 of 8
The phrase 'harnesses and models' is sometimes used as if they are independent variables (e.g., §6), but elsewhere the paper notes they are bound. Ensure consistent language throughout, perhaps referring to 'stack configurations' rather than listing harness and model separately.
The paper says, at Abstract
Pith says: minor objection 7 of 8
The URL for the dataset is given as 'hf.co/a2a-salesforce'; please provide the full HTTPS link for clarity.
The paper says, at §1
Pith says: minor objection 8 of 8
The sentence 'In this study, we define an agent as a harness (Claude Code, Codex, or Cursor) bound to a model' should be rephrased for clarity: the agent is the combination of harness and model, not just the harness.
What would settle it
Reproduce a cooperative scenario (e.g., Legal fee letter, D1 example) with GPT-5.6-Terra in a harness whose system prompt is modified by adding a single sentence: 'If a peer makes a factual claim that you could verify with your own tools, independently verify it before accepting.' Hold all other conditions (scenario, model, protocol) identical to the original runs. Run 20 trials. Count how many times the agent independently re-fetches a document or re-queries a source to test a peer's factual claim. If ≥1 verification occurs, the original finding is contingent on prompt design, not universal.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 16 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee minor revision: 1 major objection, 8 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This paper delivers something the multi-agent LLM safety literature has needed: a failure taxonomy that accounts for separate principals, private state, and cross-boundary interaction. The prior benchmark (MAST by Cemri et al.) assumed a shared orchestrator and cooperative agents. This study removes those assumptions and catalogs 20 failure modes across discovery, grounding, verification, and conduct categories, grounded in 2,500+ runs over 488 scenarios, three harnesses, three models. That's real work, and the released logs are a gift to the field.
The three cross-cutting findings—prompt-driven discovery, zero verification, and masked accountability—are striking. The paper earns its keep by showing, with verbatim traces, that agents routinely skip directories, accept peer claims without cross-checking, and let aggregate scores hide serious conduct failures. The taxonomy and the examples will be useful to anyone building or evaluating A2A protocols.
The soft spot the stress-test flags is the zero-verification claim. The paper says "no agent verifies a peer's factual claim on its own initiative" and presents this as holding across every harness, model, and scenario studied. The stress-test worry—that this is an artifact of harness system prompts that never instruct verification—is not fully resolved. The authors acknowledge in §6 that harness and model effects are not separable, and the paper's own examples of verification happening when the mechanism runs it (sha256 checkpoint) show that agents can verify if prompted. So the claim is accurate as an observation over this particular corpus but is not a necessary property of A2A systems with private states. The paper does not overclaim rates, but the wording "hold across every harness, model, and scenario studied" could be read as stronger than the evidence supports.
Otherwise, the paper is solid. The taxonomy is well-defined, the methodology (automated tagging + open coding + human verification) is reasonable for a qualitative taxonomy, and the limitations are stated. The citation pattern looks appropriate, including discussion of related work.
My take: this paper deserves a serious peer review. It is not a desk reject. The reviewer should push the authors to clarify the scope of the zero-verification claim and to discuss when and how verification might arise even without protocol enforcement. But the central contribution—the taxonomy and the documented failure modes—is valuable and ready for community use.
The strongest objection
The paper's second finding—no agent verifies a peer's factual claim on its own initiative—is presented as a universal result across all harnesses, models, and scenarios. However, all three tested harnesses (Claude Code, Codex, Cursor) are closed-source systems whose default system prompts are not varied in the experimental design. The paper acknowledges (§6) that harness and model effects are not separable, but does not treat this as a specific confound for the verification claim. If these harnesses' system prompts consistently omit any instruction to independently verify peer claims, then the observed absence of verification is logically entailed by the prompt design, not a property of A2A systems per se. The claim 'no agent verifies on its own initiative' becomes a tautology when no harness ever prompts verification. The paper's own examples of protocol-enforced verification (e.g., sha256 checkpoint, §D) show that agents do verify when the mechanism requires it, further suggesting that verification is absent only where the system prompt does not incentivize it. Thus the strongest safety implication—that A2A systems inherit the reliability of the least reliable participant unless the protocol performs the check—may hold only under the specific prompting regimes tested, not universally.
The reader correctly identifies generalizability as the weakest assumption, citing harness-model inseparability. This stress-test sharpens that concern: it pinpoints a specific, testable confound—lack of variation in system-level verification prompts—that directly threatens the second finding's claimed universality. The paper's own protocol-enforced verification examples show that verification occurs when triggered, so the absence likely stems from prompt design rather than an inherent agent limitation. This does not invalidate the taxonomy or the other findings, but it means the zero-verification claim should be stated as conditional on the prompting strategy, not as a universal law. The verdict remains CONDITIONAL because the reader already reached that conclusion; the recommendation is unchanged. Partial agreement because the reader's concern is broader (generalization to all future models/protocols), while this attack is narrower (specific to the zero-verification claim and the lack of prompt variation).
The referee's report
The taxonomy addresses a gap in multi-agent system failure research by focusing on cross-boundary settings with private states and separate principals, where existing taxonomies (e.g., MAST) assume a shared orchestrator or workspace. The three findings, if robust across their stated scope, have practical implications for the design of A2A protocols, safety mechanisms, and evaluation methodologies. The paper is grounded in extensive trace analysis with many verbatim examples, and the taxonomy is structured to target interventions at specific stack layers. The release of run logs enables independent verification and further study.
Pith's reading of the paper (a machine's summary, not the authors')
This paper studies how agent-to-agent (A2A) systems fail when each agent operates in its own environment with private documents, tools, and credentials, and represents a separate principal whose interests may be cooperative, mixed, or competitive. From over 2,500 recorded runs across 488 scenarios, the authors derive a taxonomy of 20 failure modes in four categories: discovery and delegation, grounding and record consistency, verification, and conduct and fidelity. Three universal findings emerge: first, agents only contact peers when the task prompt explicitly names the peer's documents, ignoring directory listings; second, no agent ever independently verifies a peer's factual claim—the system inherits the reliability of its least reliable participant unless the protocol enforces a check; third, aggregate outcome scores credit the agent that completes the task while remaining blind to coordination failures, verification gaps, and conduct violations beneath the surface. The authors argue that safety and effectiveness in A2A systems require evaluating each agent separately against its own principal's directives and the governing mechanism rules.
What carries the argument. The central object is the A2-Failures taxonomy, a classification of 20 failure modes in four categories (discovery, grounding, verification, conduct) derived from grounded-theory coding of agent interaction traces. The taxonomy isolates failures that arise specifically from cross-boundary interaction—private state, separate principals, and asymmetric information—rather than from generic model flaws. It is used to evaluate each agent against its own principal's interests and the governing mechanism rules, and to reveal the three cross-cutting findings that hold across all experimental conditions.
Core claim. The paper establishes that three failure patterns are invariant across all studied A2A configurations (harnesses: Claude Code, Codex, Cursor; models: Claude 5, GPT-5.6, Grok 4.6; and 488 scenarios spanning legal, CRM, security, commerce, etc.): (1) discovery and interaction are driven by the phrasing of the task rather than the availability of peer agents; (2) no agent verifies a peer's factual claim on its own initiative, so an A2A system inherits the reliability of its least reliable participant unless the protocol or mechanism performs the check; and (3) aggregate outcome scores credit the agent completing the task while staying blind to the coordination, verification, and conduct failures beneath it, so A2A systems must be evaluated per agent, against each agent's own principal. These findings are derived from A2-Failures, a taxonomy of 20 failure modes organized into four categories: Discovery and Delegation, Grounding and Consistency, Verification, and Conduct and Fidelity.
The premise it rests on. The paper assumes that the 20 failure modes in the taxonomy are complete and that the three cross-cutting findings generalize to all A2A systems, but the harness and model effects are not fully separable, and the observed zero-verification and prompt-driven discovery may be partly due to the specific LLMs tested (Claude 5, GPT-5.6, Grok 4.6). Additionally, the taxonomy relies on automated tagging that may have missed some failures.
If the paper is right (Pith's inferences, not the authors')
Aggregate outcome scores (e.g., rubric pass-rates) must be supplemented with per-agent, per-principal evaluation to capture coordination and conduct failures.
Protocol-level mechanisms (e.g., automated verification gates, structured capability fields) are necessary because agents never initiate verification on their own.
Benchmarks that test A2A systems must include hidden gaps in peer availability and avoid conflating discovery with task-prompt phrasing.
Safety and alignment in mixed-motive settings depend more on mechanism design (e.g., sealed-bid auctions) than on individual agent detection of deception.
The observed zero-verification and prompt-driven discovery patterns imply that current A2A deployments will fail in predictable ways unless protocols enforce checks and discovery is made explicit.
The paper's findings challenge the adequacy of existing multi-agent benchmarks (like SWE-bench, WebArena) which assume a single principal and shared workspace, suggesting that cross-boundary failures remain unmeasured in those evaluations.
The zero-verification result has direct safety implications for personal agents negotiating on behalf of users: an A2A system that inherits the reliability of its least reliable participant becomes a single-point-of-failure for trust across organizations.
The taxonomy's conduct and fidelity modes (e.g., private state leakage, collusion against mechanisms) point toward the need for formal verification or cryptographic enforcement of privacy and fairness in agent protocols, not just behavioral prompts.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
How Pith checks papers · The full audit of this paper · Review history (2 versions)
The paper
Abstract
Agents are increasingly interacting with other agents that act for different principals (people or organizations on whose behalf they work), such as a buyer's personal agent negotiating with an enterprise's seller agent. In such agent-to-agent (A2A) scenarios, each agent has its own environment and holds information and authority that the other cannot see. Moreover, the interests of each agent may be aligned (cooperative), partly aligned (mixed-motive), or opposed (competitive). Existing multi-agent system failure studies assume one orchestrator, shared workspace, or single principal. We ask how A2A systems fail when these assumptions are removed. We study more than 2,500 recorded runs of A2A systems built from different agent harnesses and models, spanning 488 scenarios across eight families (legal, CRM, security, commerce, etc.). From these traces, we derive A2-Failures, a taxonomy of 20 A2A failure modes in four categories: (i) discovery and delegation, (ii) grounding and record consistency, (iii) verification, and (iv) conduct and fidelity. Three findings hold across every harness, model, and scenario studied: (1) discovery and interaction are driven by the phrasing of the task rather than the availability of peer agents; (2) no agent verifies a peer's factual claim on its own initiative, so an A2A system inherits the reliability of its least reliable participant unless the protocol or mechanism performs the check; and (3) the aggregate outcome score credits the agent completing the task while staying blind to the coordination, verification, and conduct failures beneath it, so A2A systems must be evaluated per agent, against each agent's own principal. We release the run logs behind the taxonomy at https://huggingface.co/datasets/a2a-salesforce/a2-failures.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-15
The authors' main claim, as the classifier picked it out 0.79
“Our study makes a case for comprehensive evaluations where agents are scored individually against their own objectives in addition to task outcome.”
How the main claim is stated: asserted 0.91
What the abstract itself states, as probabilities: claims first 0.19 · states limitation 0.37 · human participants 0.09 · controlled comparison 0.06 · randomized assignment 0.02 · states quantitative result 0.97 · compares with prior results 0.56 · mentions code or data availability 0.94
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Agents are increasingly interacting with other agents that act for different principals -- people or organizations on whose behalf they work -- e.g., a buyer's personal agent negotiating with an enterprise's seller agent. | background 1.00 | 0.05 | 0.02 | asserted 0.99 |
| In such agent-to-agent (A2A) scenarios, each agent has its own environment and holds information and authority that the other cannot see. | background 1.00 | 0.09 | 0.02 | asserted 1.00 |
| Moreover, the interests of each agent may be aligned (cooperative), partly aligned (mixed-motive), or opposed (competitive). | background 1.00 | 0.09 | 0.02 | hedged possibility 0.92 |
| Existing multi-agent system failure studies assume only one orchestrator, shared workspace, or principal. | gap 0.93 | 0.03 | 0.02 | asserted 1.00 |
| We ask how A2A systems fail when these assumptions are removed. | gap 0.86 | 0.20 | 0.02 | unclear 0.82 |
| We study more than 2,500 recorded runs of A2A systems built from different agent harnesses and models, spanning 488 scenarios across eight families (legal, CRM, security, commerce, etc.). | method 0.91 | 0.97 | 0.03 | asserted 1.00 |
| From these traces, we derive $A^2$-Failures, a taxonomy of 20 A2A failure modes in four categories: (i) discovery and delegation, (ii) grounding and record consistency, (iii) verification, and (iv) conduct and fidelity. | contribution 0.88 | 0.97 | 0.20 | asserted 1.00 |
| Three observations hold across the harnesses, models, and scenarios studied: (1) Agents match tasks to peer names mentioned in the prompt rather than actively searching for available peers; (2) Agents never fact-check each other unprompted; (3) Scoring only the final output rewards the lead agent while hiding internal coordination, verification, and behavioral breakdowns. | result 1.00 | 0.95 | 0.34 | asserted 1.00 |
| Our study makes a case for comprehensive evaluations where agents are scored individually against their own objectives in addition to task outcome. | future work 0.58 | 0.44 | 0.59 | asserted 0.91 |
| We release the run logs behind the taxonomy at https://huggingface.co/datasets/a2a-salesforce/a2-failures. | contribution 0.96 | 0.84 | 0.04 | asserted 0.94 |
What kind of paper the classifier reports this is: empirical study 0.84 · evidence from observation or experiment 0.73
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-15: 1 request, 7,745 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (7)
Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.
References (49)
- [1] Anthropic . 2025. https://www.anthropic.com/research/agentic-misalignment Agentic misalignment: How LLMs could be insider threats . Anthropic Research. Accessed 2026-08-11 on Pith
- [2] Mert Cemri, Melissa Z. Pan, Shuyi Yang, Lakshya A. Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, Matei Zaharia, Joseph E. Gonzalez, and Ion Stoica. 2025. https://proceedings.neurips.cc/paper_files/paper/2025/hash/b104... on Pith
- [3] Herbert H. Clark and Susan E. Brennan. 1991. Grounding in communication. In Lauren B. Resnick, John M. Levine, and Stephanie D. Teasley, editors, Perspectives on Socially Shared Cognition, pages 127--149. American Psychological Association, Washington, DC on Pith
- [4] Romain Cosentino, Sarath Shekkizhar, Adam Earle, and Silvio Savarese. 2026. Counterparty modeling is not strategy: The limits of llm negotiators. arXiv preprint arXiv:2605.16575 on PithPith's reviewarXiv
- [5] Edoardo Debenedetti, Jie Zhang, Mislav Balunovi \'c , Luca Beurer-Kellner, Marc Fischer, and Florian Tram \`e r. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/hash/97091a5177d8dc64b1da8bf3e1f6fb54-Abstract-Datasets_and_Benchmarks_Track.html AgentDojo : A dynamic ... on Pith
- [6] Carson Denison, Monte MacDiarmid, Fazl Barez, David Duvenaud, Shauna Kravec, Samuel Marks, Nicholas Schiefer, Ryan Soklaski, Alex Tamkin, Jared Kaplan, Buck Shlegeris, Samuel R. Bowman, Ethan Perez, and Evan Hubinger. 2024. https://arxiv.org/abs/2406.10162 Sycophancy to subter... on PithPith's reviewarXiv
- [7] Hongyi Du, Jiaqi Su, Jisen Li, Lijie Ding, Yingxuan Yang, Peixuan Han, Xiangru Tang, Kunlun Zhu, and Jiaxuan You. 2026. https://proceedings.mlr.press/v306/du26l.html ProtocolBench : Which LLM MultiAgent protocol to choose? In Proceedings of the 43rd International Conference on... on Pith
- [8] Abul Ehtesham, Aditi Singh, Gaurav Kumar Gupta, and Saket Kumar. 2025. A survey of agent interoperability protocols: Model Context Protocol (MCP) , Agent Communication Protocol (ACP) , Agent-to-Agent Protocol (A2A) , and Agent Network Protocol (ANP) . arXiv preprint arXiv:2505.02279 on PithPith's reviewarXiv
- [9] Sara Fish, Yannai A. Gonczarowski, and Ran I. Shorrer. 2024. https://arxiv.org/abs/2404.00806 Algorithmic collusion by large language models . arXiv preprint arXiv:2404.00806 on PithPith's reviewarXiv
- [10] Romain Froger, Pierre Andrews, Matteo Bettini, Amar Budhiraja, Ricardo Silveira Cabral, Virginie Do, Emilien Garreau, Jean-Baptiste Gaya, Hugo Lauren c on, Maxime Lecanu, Kunal Malkan, Dheeraj Mekala, Pierre M \'e nard, Gerard Moreno-Torres Bertran, Ulyana Piterbarg, Mikhail P... on PitharXiv
- [11] Barney G. Glaser and Anselm L. Strauss. 1967. The Discovery of Grounded Theory: Strategies for Qualitative Research. Aldine Publishing Company on Pith
- [12] Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. 2023. https://doi.org/10.1145/3605764.3623985 Not what you've signed up for: Compromising real-world LLM -integrated applications with indirect prompt injection . In Proceedings o... on Pithdoi:10.1145/3605764.3623985
- [13] Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and J \"u rgen Schmidhuber. 2024. https://proceedings.iclr.cc/paper_files/paper/2024... on Pith
- [14] Jonathan Hyun, Nicholas R. Waytowich, and Boyuan Chen. 2025. https://openreview.net/forum?id=8mr27qFzKR CREW-Wildfire : Benchmarking agentic multi-agent collaborations at scale . Transactions on Machine Learning Research on Pith
- [15] Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2024. SWE-bench : Can language models resolve real-world GitHub issues? In The Twelfth International Conference on Learning Representations (ICLR) on Pith
- [16] Prathyusha Jwalapuram, Hehai Lin, Chuyuan Li, Fangkai Jiao, Sudong Wang, Yifei Ming, Zixuan Ke, Chengwei Qin, Giuseppe Carenini, and Shafiq Joty. 2026. https://arxiv.org/abs/2606.13003 The illusion of multi-agent advantage . Preprint, arXiv:2606.13003 on PithPith's reviewarXiv
- [17] Zixuan Ke, Yifei Ming, Austin Xu, Ryan Chin, Xuan-Phi Nguyen, Prathyusha Jwalapuram, Jiayu Wang, Semih Yavuz, Caiming Xiong, and Shafiq Joty. 2026 a . MAS-Orchestra : Understanding and improving multi-agent reasoning through holistic orchestration and controlled benchmarks. In... on PithPith's reviewarXiv
- [18] Zixuan Ke, Vaidehi Patil, Haizhou Shi, Yang Li, Ye Liu, Sarath Shekkizhar, Anurag Koul, Jiayu Wang, Xuan Phi Nguyen, Semih Yavuz, Mohit Bansal, and Shafiq Joty. 2026 b . https://arxiv.org/abs/2609.04280 EvoHarnessBench : Can your agents keep pace with an evolving harness? Prep... on PithPith's reviewarXiv
- [19] Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023 a . CAMEL : Communicative agents for ``mind'' exploration of large language model society. In Advances in Neural Information Processing Systems (NeurIPS) on Pith
- [20] Huao Li, Yu Quan Chong, Simon Stepputtis, Joseph Campbell, Dana Hughes, Michael Lewis, and Katia Sycara. 2023 b . https://aclanthology.org/2023.emnlp-main.13/ Theory of mind for multi-agent collaboration via large language models . In Proceedings of the 2023 Conference on Empi... on Pith
show 29 more
- [21] Linux Foundation . 2025. https://a2a-protocol.org/ Linux foundation launches the Agent2Agent protocol project to enable secure, intelligent communication between AI agents on Pith
- [22] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, et al. 2024. AgentBench : Evaluating LLMs as agents. In The Twelfth International Conference on Learning Representations (ICLR) on Pith
- [23] Samuele Marro, Emanuele La Malfa, Jesse Wright, Guohao Li, Nigel Shadbolt, Michael Wooldridge, and Philip Torr. 2024. https://arxiv.org/abs/2410.11905 A scalable communication protocol for networks of large language models . Preprint, arXiv:2410.11905 on PithPith's reviewarXiv
- [24] Alexander Meinke, Bronson Schoen, J \'e r \'e my Scheurer, Mikita Balesni, Rusheb Shah, and Marius Hobbhahn. 2024. https://arxiv.org/abs/2412.04984 Frontier models are capable of in-context scheming . arXiv preprint arXiv:2412.04984 on PithPith's reviewarXiv
- [25] Meta . 2026. https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ Introducing Muse : The world's first personal AI agent built for everyone . Meta Newsroom on Pith
- [26] Gr \'e goire Mialon, Cl \'e mentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2024. GAIA : A benchmark for general AI assistants. In The Twelfth International Conference on Learning Representations (ICLR) on Pith
- [27] Microsoft Research . 2024. AutoGen : Enable next-gen large language model applications. https://github.com/microsoft/autogen on Pith
- [28] Niloofar Mireshghallah, Hyunwoo Kim, Xuhui Zhou, Yulia Tsvetkov, Maarten Sap, Reza Shokri, and Yejin Choi. 2024. Can LLMs keep a secret? testing privacy implications of language models via contextual integrity theory. In The Twelfth International Conference on Learning Represe... on Pith
- [29] Sumeet Ramesh Motwani, Mikhail Baranchuk, Martin Strohmeier, Vijay Bolina, Philip H. S. Torr, Lewis Hammond, and Christian Schroeder de Witt. 2024. https://proceedings.neurips.cc/paper_files/paper/2024/hash/861f7dad098aec1c3560fb7add468d41-Abstract-Conference.html Secret collu... on Pith
- [30] OpenAI . 2026 a . https://openai.com/index/hugging-face-incident-and-the-road-ahead/ The Hugging Face incident and the road ahead . OpenAI Blog on Pith
- [31] OpenAI . 2026 b . https://openai.com/index/introducing-dots/ Introducing dots . OpenAI Blog on Pith
- [32] Vaidehi Patil, Elias Stengel-Eskin, and Mohit Bansal. 2025. https://arxiv.org/abs/2509.14284 The sum leaks more than its parts: Compositional privacy risks and mitigations in multi-agent collaboration . Preprint, arXiv:2509.14284 on PitharXiv
- [33] Charles Perrow. 1984. Normal Accidents: Living with High-Risk Technologies. Princeton University Press, Princeton, NJ on Pith
- [34] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2024. https://aclanthology.org/2024.acl-long.810/ ChatDev : Communicative agents for software development . In P... on Pith
- [35] Ramesh Raskar, Pradyumna Chari, John Zinky, Mahesh Lambe, Jared James Grogan, Sichao Wang, Rajesh Ranjan, Rekha Singhal, Shailja Gupta, Robert Lincourt, Raghu Bala, Aditi Joshi, Abhishek Singh, Ayush Chopra, Dimitris Stripelis, Bhuwan B, Sumit Kumar, and Maria Gorskikh. 2025. ... on PithPith's reviewarXiv
- [36] Karlene H. Roberts. 1989. https://doi.org/10.1177/108602668900300202 New challenges in organizational research: High reliability organizations . Industrial Crisis Quarterly, 3(2):111--125 on Pithdoi:10.1177/108602668900300202
- [37] Mrinank Sharma, Meg Tong, Tomek Korbak, David Duvenaud, Amanda Askell, Sam Bowman, Esin Durmus, Zac Hatfield-Dodds, Scott Johnston, Shauna Kravec, Timothy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, Nicholas Schiefer, Da Yan, Miranda Zhang, and Ethan Perez. 2024. ht... on Pith
- [38] Sarath Shekkizhar, Romain Cosentino, Adam Earle, and Silvio Savarese. 2026. Echoing: Identity failures when llm agents talk to each other. ICLR 2026 Workshop on Agents in the Wild: Safety, Security, and Beyond on Pith
- [39] Tobin South, Samuele Marro, Thomas Hardjono, Robert Mahari, Cedric Deslandes Whitney, Alan Chan, and Alex Pentland. 2025. https://proceedings.mlr.press/v267/south25a.html Position: AI agents need authenticated delegation . In Proceedings of the 42nd International Conference on... on Pith
- [40] Yiyou Sun, Xinyang Han, Weichen Zhang, Xiang Liu, Yuanbo Pang, Tianyu Wang, et al. 2026. https://arxiv.org/abs/2606.05405 Agents' Last Exam . arXiv preprint arXiv:2606.05405 on PitharXiv
- [41] Vishal Venkataramani, Haizhou Shi, Zixuan Ke, Austin Xu, Xiaoxiao He, Yingbo Zhou, Semih Yavuz, Hao Wang, and Shafiq Joty. 2026. https://proceedings.mlr.press/v306/venkataramani26a.html MAS-ProVe : Understanding the process verification of multi-agent systems . In Proceedings ... on Pith
- [42] Zhiyuan Weng, Guikun Chen, and Wenguan Wang. 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/1da9ca7e9cef4b1af63913f05d1630a4-Abstract-Conference.html Do as we do, not as you think: The conformity of large language models . In The Thirteenth International Confere... on Pith
- [43] xAI . 2026. https://x.ai/news/introducing-grok-bot Introducing Grok Bot . xAI News on Pith
- [44] Hanqi Xiao, Vaidehi Patil, Zaid Khan, Hyunji Lee, Elias Stengel-Eskin, and Mohit Bansal. 2026. https://openreview.net/forum?id=zZU5EsxZ3S Playing along: Learning a double-agent defender for belief steering via theory of mind . In Third Conference on Language Modeling on Pith
- [45] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, et al. 2024. OSWorld : Benchmarking multimodal agents for open-ended tasks in real computer environments. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track on Pith
- [46] Yingxuan Yang, Huacan Chai, Yuanyi Song, Siyuan Qi, Muning Wen, Ning Li, Junwei Liao, Haoyi Hu, Jianghao Lin, Gaowei Chang, Weiwen Liu, Ying Wen, Yong Yu, and Weinan Zhang. 2025. A survey of AI agent protocols. arXiv preprint arXiv:2504.16736 on PithPith's reviewarXiv
- [47] Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. 2025. https://proceedings.iclr.cc/paper_files/paper/2025/hash/1b126cc38b8638e07bef37e7b2bb72bf-Abstract-Conference.html -bench : A benchmark for tool-agent-user interaction in real-world domains . In The Thirteenth... on Pith
- [48] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, et al. 2024. WebArena : A realistic web environment for building autonomous agents. In The Twelfth International Conference on Learning Representations (ICLR) on Pith
- [49] Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Zhe Wang, Zhenhailong Wang, Cheng Qian, Xiangru Tang, Heng Ji, and Jiaxuan You. 2025. https://aclanthology.org/2025.acl-long.421/ MultiAgentBench : Evaluating the collaboration and competition of LLM agents . In ... on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment