AI agent cooperation Instinct
NP-Bench, execution consistency and earning the right to a team
Today's strongest finds separate integration success, evidence sufficiency and partner understanding from cooperation itself. These are newly found October papers unless marked catch-up; preprint results below are authors' claims, not independent replications.
Research
NP-Bench: verify the merged team result
Compares no coordination, reactive warnings and proactive scope/merge scheduling using real git merges and tests. The planner lifts deterministic clean integration from 1/9 to 9/9 scenarios; live contract-change tests also improve.
Assessment
Scope compliance, contract adaptation and integration are distinct from individual patch correctness.
Execution consistency: correct memory is not enough
October 6 paper defines duties for state use, handoffs and final-state agreement. Removing receipt/use/response evidence makes 82.4% of opposite-label execution pairs indistinguishable; restoring it separates 97.9% of the merged pairs. CAVERT diagnoses and repairs failures across benchmark settings.
Assessment
The distinction is whether information arrived and affected action, rather than merely being recorded.
Partner knowledge versus cooperative action
Across eight LLMs in a Hanabi-derived environment, probes can recover a partner's communication convention without the receiver acting on it. Giving translated action recommendations helps more than stating the convention as a rule.
Assessment
Convention inference, behavioral adaptation and team score measure different aspects of cooperation.
EvalResearchBench: test the evaluator, not just the agent
Agents design compact evaluations under fixed time/API budgets, then freeze them before testing candidate models against 14 target benchmarks. Best evaluators reproduce about 75% of pairwise rankings versus a 91% target-agreement ceiling; the development winner is not the sealed-target winner.
Assessment
Useful for validating a cooperation metric against held-out outcomes and cost.
Apple/EPFL: earn the right to add a team
Catch-up: September 30 paper, newly covered October 7. With the same strong backbone and time budget, elaborate ML-engineering harnesses offer no advantage over a minimal coding-agent session.
Assessment
A strong warning to include a budget-matched single-agent baseline.
Hermes Index: cost and contamination beside score
Nous's live leaderboard averages four suites inside one Hermes Agent harness and reports score plus mean cost per task. Its 150-task in-house suite grades files/state, often with hybrid judges. Especially useful disclosure: some SkillsBench runs found public solutions; clean scores drop those tasks.
Coordination and protocols
dovetail: verified protocols for agents with different interests
October 7 research-program essay from IDAI's Samuele Marro proposes LLM-generated interaction protocols with mechanically checked Lean proofs, covering properties such as payment, confidentiality and timelines. Plans include adversarial games and comparisons against human-designed mechanisms.
Assessment
Directly relevant to mixed-motive cooperation and protocol-level evaluation.
A2A security: check both hops
October 7 reference deployment distinguishes incoming agent calls from outgoing MCP tool calls: Cloud Run IAM protects one, network topology plus gateway policy protects the other. It also flags absent end-to-end propagation of the original caller's identity.
Assessment
Useful trust-boundary test cases for cross-owner cooperation.
Also seen
HARP: hidden preferences as numeric beliefs
October 6 paper keeps a Bayesian posterior over each agent's discrete persona rather than a prose belief, with an exploration bonus for distinguishing preferences. Tests span games, compact coordination tasks and ride-hailing dispatch.
Assessment
Relevant to cooperation among agents with conflicting goals.
DeOrch: transfer planning to an unseen worker pool
Separates worker-agnostic decomposition from worker selection, using probe-based suitability and an online contextual-bandit matcher. Reports transfer to a new pool without retraining and fewer worker calls than competing learned orchestrators.
Assessment
Useful for testing whether cooperation is pool-specific.
WorkflowOps: learn collaboration priors from past work
October 6 framework uses pairwise collaboration frequencies to guide dependency-valid workflows, adds agents when capabilities are missing, and uses embeddings before LLM routing. Reports over 80% fewer LLM routing calls.
DHCG: vary the team while it works
October 6 framework grows a hierarchical graph through planner, worker and generator roles, selectively routing information and stopping or expanding from feedback. Reports gains across code, math and domain reasoning.
Assessment
Relevant dynamic-team baseline; compare outcome gains with extra coordination and training cost, rather than treating score alone as cooperation quality.
S1-MAS: cheap control, strong reasoning
Assigns bounded routing, inspection and stopping decisions to lightweight models while stronger workers do open-ended reasoning. Seven-benchmark experiments report better accuracy with lower inference cost.
Assessment
This separates control cost from worker reasoning.
PIP: coordinate when the partner disappears from view
October 6 embodied zero-shot coordination work predicts hidden partner location and behavior from local history, with richer joint observations during training. Evaluates Burrito-PO, Overcooked-PO and Melting Pot plus a human study.
Assessment
The study examines coordination with unseen partners and occlusion.
CASTLE: counterfactual guidance without communication
October 6 decentralized MARL framework trains local-dynamics and semantic-social world models from counterfactual simulator rollouts, then uses frozen models with local information at execution.
Assessment
Relevant as a communication-free cooperation baseline.
LiRA: who bears a shared constraint?
Learns responsibility shares for a common cost budget, reporting up to 29% better welfare across four environments with 3-400 agents.
Assessment
Relevant to cooperation under unequal burdens: measure welfare and constraint violations together.
Atlassian MCP: broader surface, scoped risk
October 6 announcement expands to 220+ tools, with toolset scopes and separate read/write/destructive risk tiers.
Assessment
Useful enterprise cooperation substrate, but vendor efficiency claims are not an independent result.
MCP Toolbox Java SDK 1.0
Google's original announcement is October 6, despite October 7 reposts. Adds transport abstraction, per-request credential resolution, version/session handling and removal of server-bound parameters from exposed tool definitions.
Assessment
Useful for tenant-boundary and identity-preservation tests.
Unverified leads
Swarm gains with stronger models: primary still missing
An October 8 social roundup points to a Prime Intellect residency write-up arguing stronger models benefit more from cooperative communication.
Editorial synthesis
Joint-outcome checks, evidence-complete handoff traces and the distinction between understanding a partner and acting on that understanding offer complementary views of cooperation. Equal-budget single-agent baselines and held-out tasks and partner pools help put team results in context.
Editorial interpretation of the cited sources.