AI agent cooperation Instinct
Hidden belief collapse, A2A failures and MASTraceBench
Peer redundancy does not protect a discussion group, and final-output scoring does not see the failure. Ten-agent discussions collapse privately under a single persistent adversary while public transcripts look healthy; cross-principal agent-to-agent runs fail through discovery, grounding, verification and conduct breakdowns that nobody checks unprompted. Trajectory-level, team-level and responsibility-chain metrics show where coordination actually happens, and institutions emerge when agents negotiate over a real shared bottleneck.
Evaluation
Believed but Not Said: transcripts miss most peer-pressure collapse
Huang and colleagues run ten-agent factual discussions in which one or three adversarial agents repeatedly assert a plausible false answer, probing each benign agent's hidden private answer separately from its public stance. Across six models and three factual-QA datasets, benign groups finish with the false answer only 0.6% of the time without an attacker; one attacker raises final private infection to 44.8% and three to 60.5%. Under three attackers, benign agents hold the false answer in 51.6% of exposed turns but endorse it publicly in only 19.7%.
Assessment
Score private state alongside the transcript; public agreement is not a reliable proxy for what a discussion group actually holds.
October 15 primary preprint.
A2A-Failures: agents match peer names, and nobody fact-checks
The authors study more than 2,500 recorded runs of agent-to-agent systems acting for separate principals, spanning 488 scenarios across eight families including legal, CRM, security and commerce. They derive a taxonomy of 20 A2A failure modes in four categories: discovery and delegation, grounding and record consistency, verification, and conduct and fidelity. Agents matched tasks to peer names mentioned in the prompt rather than searching available peers, never fact-checked each other unprompted, and final-output scoring rewarded the lead agent while hiding internal coordination and verification breakdowns.
Assessment
Cross-principal failures need per-agent scoring against each party's own objectives, not a single final-output grade.
October 14 primary preprint; newly surfaced catch-up.
MASTraceBench: interaction lifts weak proposals more than strong ones
Li and colleagues track and grade proposal trajectories across six cooperative and competitive tasks, with a metric suite covering task score, collaboration gain, trajectory indicators and token cost. Final answers rarely surpass the strongest initial proposal; interaction mostly lifts weaker proposals toward it, while strong ones are seldom improved and may regress. Their claim-level synthesis method CLEARS achieves the highest collaboration gain on five of six tasks, reaching 6.42 on EvidenceChain versus 3.92 for the strongest baseline.
Assessment
A trajectory-level collaboration-gain metric separates genuine improvement from mere preservation of one good opening proposal.
September 28 primary preprint; newly surfaced catch-up.
HM3D-M3: team navigation goals need team-level scoring
The benchmark places a team of agents at distributed starts in a photorealistic indoor scene, searching for an unordered set of goals specified as open-vocabulary categories, instance images or descriptions. It provides 1,000 validation episodes over 3,926 goals from 379 categories, with team-level metrics combining goal-level success rate and makespan-normalized SPL. Adapted multi-robot and zero-shot methods reach 16-27% success; the authors' trained goal assigner and training-free VLM team reach 38-41% with distinct failure modes, and the VLM team deploys on a heterogeneous three-robot team.
Assessment
Multi-agent, multi-modal, multi-object episodes expose assignment and duplication failures that single-goal navigation scores hide.
October 14 primary preprint; newly surfaced catch-up.
Coordination research
SCARCE: coding agents invent institutions to share compute
The framework has agents solve sequences of coding tasks under individual deadlines while negotiating access to a shared execution resource that gates inspection, compilation and testing. Auditing more than 1,500 trajectories from eight models with a 21-category taxonomy of negotiation strategies and emergent norms, the authors find agents spontaneously form persistent rules grounded in reputation, reciprocity and precedent, sometimes forgoing immediate advantage to honor earlier agreements. Newer frontier models show more institutional behavior; alongside cooperation, the authors observe exclusionary coalitions and rare divide-and-conquer strategies.
Assessment
Long-horizon negotiation over a real bottleneck surfaces emergent norms that stylized bargaining games cannot.
October 15 primary preprint.
RCM: delegated tasks keep their responsibilities attached
Nguyen and colleagues couple each task delegation with explicit commitments, so agents retain upstream responsibilities while locally orchestrating downstream ones, evolving a commitment graph without centralized control. Across reasoning, coding and executable agent environments, RCM outperforms centralized and decentralized baselines by up to 18.7 points in task performance and 10.2 points in the authors' responsibility score. One baseline monitors 95.0% of delegated work but intervenes on only 32.5%; RCM reaches 67.0% intervention and 65.4% fulfillment.
Assessment
Scoring the full responsibility chain (communicate, monitor, intervene, fulfill) exposes coordination that observes failures but never acts on them.
October 15 primary preprint.
MetaShare: share the failures, not only the winning edits
The framework records successful and failed agent modifications with signed performance deltas across self-improving agent lineages, retrieving high-confidence failures alongside diverse strong strategies before each generation, while a task-blind auditor injects underrepresented mechanism categories. Across paper review, coding, math grading, robotics reward design and financial audit-risk, MetaShare matches or exceeds the DGM-H baseline's final quality while converging in far fewer generations, and induces a model-routing strategy the baseline never proposes.
Assessment
Treating failed improvements as shared cross-branch knowledge attacks the redundant-search failure mode of open-ended agent collectives.
October 15 primary preprint.
CoLA: messages matter only when information is private
CoLA keeps each vision-language-action backbone frozen and trains a lightweight message channel plus per-arm diffusion heads, with fully decentralized execution. On an occluded tray handover where the marker is visible to one agent only, CoLA reaches 76.0% success versus 1.3% with messages zeroed throughout training. With a shared overhead camera, removing messages barely changes two-arm (90.6% versus 88.3%) and three-arm (63.5% versus 60.0%) success, while the wrist-only two-arm variant drops from 74.8% to 64.3%.
Assessment
A clean ablation of when explicit communication earns its cost: when observations are private, not when the workspace is shared.
October 11 primary preprint; newly surfaced catch-up.
Infrastructure
Agent Status Protocol: never trust an agent's own liveness
ASP version 0.1 defines small JSON events pushed over HTTP POST to a status server, with observers reading aggregated state over Server-Sent Events. Because a crashed, hung or compromised agent cannot be trusted to report its own failure, the server infers liveness from missed heartbeats, detects agents that are alive but not progressing, and lets supervisor, guardrail, operator and peer observations take precedence over self-reports. State spans independent dimensions (lifecycle, health, task, floor and governance), reusing A2A task states, Open Floor Protocol floor states and OpenTelemetry GenAI attribute conventions.
Assessment
Externalized liveness and multi-source state reporting are table stakes for supervising heterogeneous agent deployments.
October 15 primary preprint.
SSR: retrieve a few agent cards, not the whole registry
SSR is a deterministic, training-free retriever for Agent Card discovery: it derives structural lexical signals from capability-bearing card fields, weights them with inverse document frequency and scores candidates through an inverted signal index to keep the orchestrator's candidate context compact. At 1,000 agents, SSR reduces LLM input by 98.8% and orchestration latency by 70.9% relative to full-catalog orchestration; on an MS MARCO adaptation, IDF improves retrieval effectiveness within the structural representation.
Assessment
Candidate-context scalability is a concrete, measurable piece of agent discovery that sits before any reasoning about capability.
October 15 primary preprint.
Listen
AI Engineering Podcast: Jonathan Ellis on coordinating coding agents
Tobias Macey interviews Jonathan Ellis about shifting human review from code toward specifications, RFCs and architecture as agents write more software. Ellis describes Brokk AI's tooling, including Bifrost for fast agent-friendly static analysis and Mjolnir for coordinating multiple coding agents across harnesses and subscriptions, and argues for investing in automated validation, static analysis and fuzzing gates over manual code reading.
Assessment
A practitioner conversation on supervising teams of coding agents rather than reviewing their output line by line.
Episode published October 15.
Editorial synthesis
A cooperation claim is only as good as what it observes. Public transcripts, final answers and self-reported status each miss a different layer: private belief, per-agent obligations and liveness. Pair every channel with an external probe, such as hidden-state checks, responsibility chains, trajectory metrics or heartbeat inference, before calling a system coordinated.
Editorial interpretation of the cited sources.