AI agent cooperation Instinct
DuoCook's 26% kitchen, mid-task recovery and AAIF's new projects
DuoCook finds one VLM controlling two kitchen agents completes barely a quarter of orders, and a journal framework scores mid-episode recovery from injected changes. Coordination work covers governed self-evolution, coupled scientific search and per-edit credit assignment.
Evaluation
DuoCook: two kitchen agents under one VLM complete barely a quarter of orders
Jiao and colleagues introduce DuoCook, an Overcooked-inspired benchmark in which one vision-language model controls two cooking agents from top-down visual observations across 20 Unreal Engine maps in five categories covering equipment contention, changing terrain and handoff dependencies. Four VLMs (GPT-6 Astra, Claude Opus 5, Qwen3.8-27B and Gemini 3.5 Flash Lite) are scored on task score, order completion, delivery precision, action acceptance rate and API cost. The strongest model, GPT-6 Astra, reaches only 26.2% mean order completion. Ablating a shared, self-authored notebook between the two players raises scores for Claude and Qwen but lowers them for GPT and Gemini. The environments themselves were built by coding agents using MCP tools and Markdown skills.
Assessment
A clean signal that centralized visual coordination of two embodied agents is far from solved, and that shared notebook memory is not reliably beneficial across model families.
October 22 preprint, announced October 23; first coverage.
A journal framework scores how agents recover when the task changes mid-episode
Balyan and Sharma pair three agent architectures (ReAct, Reflexion and Plan-and-Execute) with three deterministic synthetic environments (tool planning, debugging and constraint planning) and inject tool denial, goal modification or constraint injection at graded severity, across two matched factorial grids of 4,320 episodes each, one with an explicit perturbation notice and one silent. ReAct attains the highest pooled reward at 0.544. Plan-and-Execute leads on tool planning at 0.711 but collapses on constraint planning at 0.114. Tool denial uniquely depresses behavioral recovery, severity shows no monotone dose-response, and telling the agent about the perturbation does not improve mean reward for any architecture.
Assessment
Static end-task accuracy hides exactly the failures this grid exposes, and the noticed-versus-silent design gives a reusable method for testing mid-episode adaptation claims.
Published October 20 in Scientific Reports, four days before this report; catch-up coverage.
Coordination research
Steward: nine trading agents improve themselves inside a core no self-edit can reach
Feng and colleagues present Steward, an autonomous investing system in which nine role-specialised agents decide jointly: four analysts read news, SEC filings, macroeconomic series and market data and emit typed reports, a Portfolio Manager sets target positions and a protected Risk Manager audits them. A seven-layer shared memory, assembled by code from the system's own record, supplies position context, market narrative, calibration and past decisions. The system grades every trade and forecast nightly, and each quarter a reflector proposes prompt learnings that a deterministic gate applies reversibly, verified by replay of recorded inputs. An invariant core of order arithmetic, risk rules and discipline-agent prompts sits outside the reach of any self-edit. In a five-month live deployment of 102 trading days and 1,632 LLM calls, the system contained a 24.6% sector drawdown to a 0.3% loss, and in replay a single model given the same raw data lost 1.35 points against the market over five days where the joint design showed no significant loss. No injected instruction was followed, and a 2015 to 2026 backtest shows the core alone lowering maximum drawdown under all 50 random entry policies.
Assessment
The typed boundary between self-evolution and an invariant core is a concrete, testable answer to how learning agents keep safeguards intact, and the live deployment plus replay design is stronger evidence than a backtest alone.
October 22 preprint, announced October 23; first coverage.
ADAPT treats scientific optimization as two coupled decisions and halves detector-design loss
Feng, Ni and colleagues frame scientific optimization as two coupled decisions, where to search and how to search. ADAPT uses historical designs to pick promising and diverse initial conditions, progressively concentrates computation on favorable directions, and evolves the optimization procedure itself from execution feedback. On gravitational wave detector design with a differentiable physics simulator and eight unseen topologies, ADAPT reaches a mean loss of 0.204 (plus or minus 0.101), 54.9% below the best classical optimizer and 62.2% below the best general-purpose agentic evolution baseline, with the lowest loss on all unseen topologies among evaluated baselines.
Assessment
Separating where to search from how to search, then letting the system adapt both, is a reusable pattern for budget-limited cooperative optimization rather than a domain trick.
- ADAPT for Science: A Multi-Agent System for Self-Evolving Scientific Optimization (arXiv 2610.30851) —
October 22 preprint, announced October 23; first coverage.
RoleFlow gives each structural edit its own counterfactual credit
Ge, Wang and colleagues address credit assignment in self-evolving multi-agent orchestration, where trajectory-level rewards assign one outcome to every decision and cannot isolate individual structural changes. RoleFlow refines collaboration structures through feedback-driven atomic edits and on-demand role synthesis, and its Single-Edit Counterfactual Policy Optimization evaluates each edit against its no-edit counterpart at the same chain state, assigning targeted credit to edit-specific tokens. Across seven benchmarks in four domains, RoleFlow posts the best average, 4.7 points above the strongest competitor, and leads all three multi-hop QA benchmarks by 3.2 to 7.5 F1 points, with reported cross-benchmark transfer and cross-executor generalization.
Assessment
Counterfactual credit at the single-edit level is a methodological step the orchestration literature needs, since most reported gains still rest on trajectory-level rewards.
October 21 preprint, announced October 22, two days before this report; catch-up coverage.
Infrastructure
Agentic AI Foundation takes in Agent Skills, Open Responses, OpenEnv and OpenSandbox
The Agentic AI Foundation announced at AGNTCon + MCPCon North America that four companies contributed four projects to neutral governance: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba. The additions cover agent capabilities, cross-provider model interoperability, training and evaluation environments, and secure execution, and bring the foundation to 10 hosted projects a year after launching with three, alongside A2A, MCP, AGENTS.md, goose, Agent Router and agentgateway.
Assessment
Four competitors putting adjacent layers of the agent stack under one vendor-neutral roof is the strongest governance consolidation signal since the foundation launched.
Announced October 22 at AGNTCon + MCPCon in San Jose, two days before this report; catch-up coverage.
Listen
AI Post Transformers: a skeptical listen on whether prompted agents form real collectives
The hosts examine Christoph Riedl's ICLR 2026 paper asking whether groups of LLM agents form genuine collectives or remain aggregates of individuals. Ten GPT-4.1 agents play a group binary search game in which each privately guesses a number so the total hits a hidden target, succeeding only by taking complementary offsets. The episode walks through partial information decomposition into unique, redundant and synergistic parts, and stresses that the results depend on the chosen redundancy definition, that the leading alternative measure is not reported, and that Fisher's method and Wilcoxon tests stand in statistical tension across the three prompt conditions.
Assessment
A useful counterweight to emergence claims: the choice of measure, not only the model behavior, decides whether coordination looks real.
Episode of October 19, five days before this report; catch-up coverage.
SE Radio 743: Sowjanya Pandruju on managing agent state
Sowjanya Pandruju, cloud application architect at AWS, speaks with host Gavin Henry about what agent state is, why agents need it, the different types of state, where to store it and how multi-agent systems change the picture, along with solutions to the challenges that shared and long-lived state creates.
Assessment
A practitioner-level treatment of the state layer that coordination protocols assume but rarely specify.
Episode of October 21, three days before this report; catch-up coverage.
Editorial synthesis
Today's evaluation items attack the same gap from different sides: what agents do after the plan meets the environment, whether that means two players sharing one kitchen or one agent absorbing a mid-task change. The coordination papers converge on governed adaptation, systems that edit their own collaboration structure only inside boundaries they cannot cross.
Editorial interpretation of the cited sources.