AI agent cooperation Instinct
Hidden profiles, Vals AI team costs and Sakana MASS
New evaluation work separates fact sharing from integration and team gains from extra compute. A versioned teamwork benchmark warns against mixing old and new scores, while Sakana MASS tests learning from team trajectories. Source results are not independently replicated; catch-up dates are explicit.
Evaluation
Hidden profiles: sharing facts is not integrating them
This study tests three-agent groups from four model families across shared versus hidden information and synchronous versus asynchronous communication. Across 219 valid sessions, two models share private facts at rates above 75% yet fail to combine them into the correct decision.
Assessment
Information transmission and collective decision quality need separate metrics.
October 10 CSCW companion paper.
Vals AI: compare teams with more reasoning, not only solo defaults
Katrina Drozdov reports 50 app-building tasks for two models, solo and in teams, at medium and maximum effort. Teams cost 1.8-5.1 times more; only Sol at medium effort shows a significant score gain, 7.3 points.
Assessment
Higher solo reasoning effort is an important budget alternative to adding agents.
October 9 primary report, newly surfaced through October 11 coverage.
JaxAHT v1.1: held-out partners and non-comparable old scores
The release adds validation and held-out teammate sets, trained best responses, human-proxy agents, Hanabi and larger foraging tasks. It also fixes recurrent-state and confidence-interval issues and retunes hyperparameters.
Assessment
Versioned correctness changes can invalidate comparisons even when the benchmark name stays the same.
Catch-up: actual release September 15, not the October 12 repost.
Station: score research coverage, not paper resemblance alone
Station adds supervision and periodic reflection to a six-agent research ecosystem. Authors report rediscovering 62.7% of predefined findings across three withheld-paper tasks, versus 15.4% for Codex Multiagent-v2 and 14.4-20.6% for AI Scientist-v2.
Assessment
Predefined sub-discoveries and supporting evidence offer a more inspectable research score.
Stay In Your Lane: deterministic scope checks for a task agent
Susmita Biswas tests ten models on 40 requests for a challenge-design JSON object. Code checks parsing, schema, faithfulness, deferral and forbidden referee fields such as scores and coordinates.
Assessment
Task understanding and staying within delegated responsibility can fail independently.
October 11 community benchmark write-up.
Coordination research
Sakana MASS: team trajectories train the shared model
Sakana trains a shared 27B model through two cycles of team execution, workflow search and self-selection. Authors report a 69.9% test-task win rate against the base and 1.2-1.6 times its score per output token on four research benchmarks.
Assessment
Coordination traces can supply training signal for both task execution and workflow design.
October 12 announcement.
Forager's Dilemma: who pays to enforce cooperation?
Ghasemi, Singh and Crowley compare four governance regimes in a shared-resource environment. Combining intrinsic priors with social enforcement yields roughly 1% lying and high sustainable gathering in their Q-learning and PPO experiments.
Assessment
Peer policing has a cost, so evaluating critique alone can miss incentives to stop enforcing norms.
October 11 proceedings listing.
Financial Town: trace dialogue into decisions
This poster separates fixed market trajectories, social information and agent cognition. Paired social/no-social runs find no stable aggregate shift but heterogeneous persona-level changes, with sparse inspectable dialogue-to-decision traces.
Assessment
Measure whether a message is retrieved and affects an action, not merely whether it was sent.
October 10 CSCW companion poster.
Protocols and infrastructure
MCP reference servers: fetch guard shipped, not a new protocol
The October 11 repository release lists mcp-server-fetch 2026.10.10 and its default refusal of private, loopback and metadata addresses. The current README also documents redirect checks.
Assessment
Adds an explicit network boundary when untrusted content influences a fetch target.
October 11 release, verified against GitHub release metadata.
6G agent requirements: proposed dialogue continuity
Sarikaya and Schott discuss agent discovery, network requirements and a proposed Agentic Dialog Management Protocol for continuity and correlation across user-agent, agent-agent and agent-tool exchanges.
Assessment
Tracks dialogue identity as infrastructure, rather than assuming each tool call is an isolated interaction.
Draft last updated October 11.
Listen
The Changelog: From zero to NanoClaw
Adam Stacoviak and Gavriel Cohen discuss container isolation, enterprise identity, audit and approval policies in an approximately 2-hour-43-minute interview.
Assessment
Worth listening for the distinction between a personal-agent sandbox and organization-wide governance.
Catch-up listen: aggregator dates it October 9, while later summaries surfaced it October 11.
Editorial synthesis
Cooperation quality needs evidence at several levels: information arrives, is integrated, stays inside authority, and improves the result at a fair budget. Version fixes and training-only gains can change the story without proving better deployed teamwork. Network guards and dialogue identity help define boundaries, but do not replace outcome checks.
Editorial interpretation of the cited sources.