AI agent cooperation Instinct
Concordia and Anthropic on what cooperation scores really measure
A quieter news cycle makes careful catch-up more useful than recycled announcements. Concordia and Anthropic lead on what cooperation scores really measure; newer work adds shared-world consistency, explicit research institutions and counterfactual handoff tests. Results remain source-reported, not independently replicated.
Evaluation
Concordia: cooperation with unfamiliar partners
Smith and colleagues evaluate mutual gains across mixed-motive social situations, including 14 held-out scenarios. Agents struggled particularly with persuasion and norm enforcement.
Assessment
A direct cooperation-evaluation template: vary partners and situations instead of measuring one familiar team on one task.
Catch-up: December 3, 2025 paper. The October 8 ACM listing is not a new experiment.
Anthropic: integration metrics can hide weak cooperation
Anthropic reports swarms building games, scanning code and facing conflicting objectives. Some models merge more PRs by avoiding shared files; coordinated vulnerability scanning also searched more code than its independent baseline.
Assessment
Merge rate, code sharing, search scope and token budget need separate controls before attributing gains to cooperation.
Catch-up: August 13 report, resurfaced in October commentary.
Rational policy gradient: avoid training self-sabotage
Lauffer and colleagues constrain adversarial partner training to policies that are optimal against some possible partner. They report stronger robustness, adaptation and diversity in cooperative and general-sum games.
Assessment
Stress-test partners can expose real weaknesses without making impossible, reward-destroying behavior the training target.
Catch-up: 2025 work. Recent ACM indexing does not make it a new method.
SMDD-Bench: isolated verification for long-horizon science
The Harbor package contains 502 small-molecule design tasks, the original harness and a GPU oracle. Agent-visible task containers and separate verifier containers isolate the submitted artifact from grading files.
Assessment
A useful pattern for measuring tool use and planning while keeping evaluators out of the agent workspace.
Catch-up benchmark with a Harbor package; exact package-release date not verified.
ME-World: measure whether viewpoints agree after an action
Chung and colleagues jointly generate embodied agents' first-person streams with shared action conditioning and environment memory. Their metrics separate environment, state-update and counterpart-identity consistency.
Assessment
Checks whether an interaction changes the same world across views, rather than scoring each video in isolation.
October 8 preprint, newly found today.
Coordination research
A Society of Researchers: institutions, not just more agents
Asaria, Gandhi and Salomone describe persistent research agents competing for compute through proposals, independent review and grants. A human governor allocates resources rather than assigning each task.
Assessment
Makes allocation, verification, reputation and diversity explicit experimental variables.
October 7 source listing; paper front matter also shows 2027, so manuscript-year metadata is unresolved.
MA-JEPA: joint prediction, decentralized execution
Kaplowitz, Obahor and Schroeder de Witt predict future observation embeddings instead of reconstructing observations. A training-only joint model informs each agent's local dynamics; the authors report strong results on eight StarCraft tasks.
Assessment
Separates information sharing during training from what an individual agent needs when acting.
Newly found paper; original submission date was not recovered.
RoboAware: compare handoff choices from the same state
Zhou and colleagues restore identical simulator states to test both modular skills and end-to-end policies, then learn a coordinator that selects a policy family at deployment. They report 77% overall success across 100 tasks.
Assessment
Counterfactual branch outcomes make responsibility changes measurable rather than judging only the chosen path.
Infrastructure lead
BAND: collaboration needs more than task calls
Vlad Luzin's 17-minute AI Engineer talk describes ordered delivery, persistent conversations, runtime mapping and permissions. Its demos connect cross-user discovery and approved contacts to owner-visible coding work.
Assessment
The missing-plumbing checklist is useful even when a protocol already handles message exchange.
Surfaced by October 10 commentary; the talk publication date is unverified.
Listen
The Cognitive Revolution: how agents decide
Nathan Labenz and Goodfire's Eric Bigelow discuss critical tokens, model beliefs, uncertainty and the limits of readable reasoning in an approximately two-hour episode.
Assessment
Most relevant publisher chapters: modeling beliefs (1:07:35), phase shifts and uncertainty (1:16:22), uninterpretable chain of thought (1:31:49) and automated research (1:45:57).
October 10 episode.
Editorial synthesis
Better cooperation is not the same as more merged code, more messages or a larger agent count. Evaluations should vary partners, control budget and search scope, check shared outcomes, and distinguish successful isolation from productive joint work. Institutions and handoff rules are hypotheses to test, not gains to assume.
Editorial interpretation of the cited sources.