Multi-agent coordination

Stanford HAI's report on CooperBench says paired coding agents can lose a large share of their solo capability. This does not establish that multi-agent systems are generally inferior. It supports a narrower claim: adding actors introduces coordination variables that must be designed and measured.

The historical literature predicts the failure modes: ambiguous task allocation, inconsistent beliefs, communication cost, strategic behavior, duplicated work, unowned integration, and correlated errors. More messages do not automatically create shared understanding.

Let useful parallel work be fraction p, serial/integration work be 1-p, and n agents operate in parallel. Amdahl's idealized speedup is S(n)=1/[(1-p)+p/n]. Add coordination overhead c(n) as a fraction of solo time:

S_coord(n)=1/[(1-p)+p/n+c(n)].

Worked example: with p=0.8, two agents, and c(2)=0.15, speedup is 1/(0.2+0.4+0.15)=1.33, not the ideal 1.67. If overhead rises to 0.45, speedup falls below one: the team is slower than one agent.

Coordination is also a game. A protocol must align local action with joint utility. Contract Net uses task announcements, bids, awards, and reports; blackboard systems coordinate through shared state; swarm systems rely on local rules. Each moves complexity rather than eliminating it.

1. 1980 — Contract Net. Reid Smith proposed distributed task allocation through a contracting protocol. IEEE Transactions on Computers (opens in a new tab).
2. 1980 — Hearsay-II. Specialist knowledge sources coordinated through a shared blackboard and scheduler. ACM Computing Surveys (opens in a new tab).
3. 1981 — Evolution of cooperation. Axelrod and Hamilton used iterated prisoner's-dilemma tournaments to study conditions under which cooperation emerges. Science (opens in a new tab).
4. 1987 — Boids. Craig Reynolds showed that simple local rules could create flocking behavior. Emergence did not require centralized planning, but the task was tightly structured. SIGGRAPH (opens in a new tab).
5. 1996 — Ant System. Dorigo, Maniezzo, and Colorni used distributed pheromone-like updates for combinatorial optimization. IEEE (opens in a new tab).
6. 1997 — RoboCup. Kitano and collaborators proposed robot soccer as a shared challenge for real-time cooperative multi-agent research. AI Magazine (opens in a new tab).
7. 2005 — Swarm-bots. The SWARM-BOTS team demonstrated self-assembling robots and exposed the gap between simulated collective behavior and physical execution. IEEE Robotics & Automation Magazine (opens in a new tab).

1. 2017 | Multi-Agent Deep Deterministic Policy Gradient. Ryan Lowe and collaborators trained decentralized actors with centralized critics in mixed cooperative and competitive environments. The method addressed non-stationarity created by learning peers, but the experiments used small simulated populations and scaling remained difficult. NeurIPS (opens in a new tab).
2. 2018 | QMIX. Tabish Rashid and collaborators factorized a centralized action-value function into monotonic per-agent values that permit decentralized execution. StarCraft II micromanagement experiments advanced credit assignment, while the monotonic constraint excluded some valid coordination structures. ICML (opens in a new tab).
3. 2019 | AlphaStar. Oriol Vinyals and collaborators combined imitation learning, reinforcement learning, and league training to reach Grandmaster level in StarCraft II. League opponents operationalized an evolutionary population idea, but the result depended on one game, human replay priors, and substantial compute. Nature (opens in a new tab).
4. 2021 | Melting Pot. Joel Leibo and collaborators introduced social-dilemma substrates and held-out test scenarios for evaluating how multi-agent populations generalize. The suite improved experimental control, but stylized games remain imperfect evidence about organizations or societies. Primary paper (opens in a new tab).
5. 2023 | Generative Agents. Joon Sung Park and collaborators combined language-model memory, reflection, and planning in a 25-agent sandbox. Human evaluations found the architecture more believable than ablations, but the study was a small scripted simulation and did not validate real social or organizational behavior. UIST (opens in a new tab).

Objective: compare one agent with two-agent designs on tasks with controllable decomposability.

Environment: 120 deterministic software-maintenance simulations or document transformations. One third decomposes cleanly, one third has shared dependencies, and one third is serial. Use the same model, tools, token budget, and evaluator.

Conditions: single agent; two independent agents plus selector; planner/executor; two specialists with explicit interface contract and integrator.

Metrics: verified success, wall-clock latency, total tokens, duplicate work, merge conflicts, messages, and recovery after one injected error.

Expected result: cleanly decomposable tasks may benefit; serial and shared-state tasks may not. Explicit interfaces should outperform unconstrained conversation when specialization is real.

Success criterion: a multi-agent design improves quality-adjusted latency on a prespecified task class without increasing total cost beyond the declared limit.

Falsification: no class shows a net benefit after equivalent budgets, or improvements vanish when the integrator's work is counted.

Pitfalls: giving teams more total compute, scoring by agent self-report, and selecting the best condition after seeing the test set.

Claim: explicit ownership and interfaces reduce multi-agent integration failures.

• Baseline: free-form agent conversation.
• Intervention: task contract containing owner, input, output schema, deadline, acceptance test, and escalation rule.
• Primary metric: verified integration success.
• Secondary metrics: duplicated actions, unresolved conflicts, cost, and latency.
• Uncertainty: paired tasks and confidence interval on the success difference.
• Stopping rule: freeze protocol and evaluator before the final 60 tasks.
• Disconfirming result: no improvement or the protocol's overhead cancels its reliability gain.

Stanford HAI (opens in a new tab), Microsoft Research (opens in a new tab), Google DeepMind (opens in a new tab), MIT CSAIL (opens in a new tab), Berkeley AI Research (opens in a new tab), Scale Labs (opens in a new tab), and robotics groups including Physical Intelligence (opens in a new tab). The immediate benchmark signal is Stanford's CooperBench report (opens in a new tab).

• High confidence, 6–18 months: multi-agent claims will face stronger single-agent, ensemble, and equal-compute baselines. Disconfirming signal: benchmark reports continue to omit these controls without pushback from users or reviewers.
• Medium confidence, 12–30 months: successful systems will use fewer, more specialized agents with explicit contracts rather than large conversational swarms. Disconfirming signal: loosely coordinated swarms show consistent net gains after cost and integration are counted.