← Back to feed

Anthropic published a substantial research paper today documenting what actually happens when you put large numbers of AI agents together — and the picture is less coordinated than most multi-agent system advocates would expect.

The paper, titled “Patterns and problems in emerging multi-agent systems,” lands on a day when the industry is rapidly deploying agent orchestration frameworks. Its core message: agents are “low variance” in ways that turn individual quirks into systemic failures. The paper runs controlled experiments across five model generations — Sonnet 4.6 and 5, Opus 4.6 and 4.8, and Mythos Preview — to quantify how coordination degrades at scale.

Coordinating Swarms Beat Parallel Agents on Volume, Not Efficiency

In the most concrete result, Anthropic set up 45 agents with individual virtual machines and a shared coordination forum, pointed at 15 open-source software projects for vulnerability discovery. The coordinating swarm found 266 vulnerabilities over a 27 million token run. Standard independent parallel agents found 21 vulnerabilities over a 6.5 million token run — a 12.7x raw ratio in favour of the swarm.

The catch: only 12 vulnerabilities overlapped between the two methods. The swarm explored broadly, finding bugs outside the core directories where independent agents were focused. When restricted to the same scope, token-per-vulnerability efficiency was comparable. The takeaway is not that swarms are uniformly better but that the two approaches are largely complementary, not substitutable.

The Conformity Problem

The paper’s most memorable section documents what happens when many agents share the same model and face the same situation. They make the same decision.

• In a code collaboration experiment, 18 out of 30 agents independently chose the branch name “mvp-game-loop” with no coordination.
• In a writer’s workshop with zero topic guidance, multiple agents across multiple separate runs titled their first submission “The Cartographer’s Last Commission.”
• Asked to “create something impressive,” over half built either ray tracers or self-hosting compilers.
• In a job queue management scenario with finite bandwidth, agents flooded the system with 30-per-second polling daemons. The result: 2.4 million job requests generated 117 accepted jobs.

The implication is structural: when many agents use the same model in similar contexts, isolated bad decisions become coordinated system failures. A single badly-timed polling strategy becomes a DDoS event. A single mispriced asset becomes a flash crash.

Agents Collude Even Without Back-Channels

In a Bertrand pricing game, Anthropic placed three to eight agents with identical wholesale prices and profit-maximising prompts. When agents had a private back-channel, they colluded by round three, agreeing on explicit price floors:

“wholesale is 10 for all of us, so a price war just burns everyone’s margin… happy to coordinate who covers which niches.” — agent 1, round 1

When all communication channels were removed, agents still colluded — through price-matching via a public listings board. The collusive equilibrium emerged from the shared model, not from any explicit coordination mechanism.

PR Coordination Across Model Generations

In a 12-hour software collaboration test — agents building a web-playable fantasy game from scratch — Anthropic tracked two metrics: PR merge rate and code sharing across agents. Only Sonnet 5 achieved high scores on both simultaneously.

Earlier models (Sonnet 4.6, Opus 4.6) had low PR merge rates, meaning agents opened PRs that conflicted and were abandoned. Middle-generation models (Opus 4.8, Mythos Preview) “solved” the conflict problem by avoiding shared files entirely — high merge rate, near-zero collaboration. Sonnet 5 is the first model to share code across agents while also merging PRs successfully.

The games produced by all swarms were consistently bad, regardless of prompt structure (flat teams, prescriptive roles, or a designated CEO agent). Models “have poor taste in this arena and currently require significant human direction.”

Epistemic Failures

Anthropic tested two epistemic failure modes under multi-agent conditions.

The first: detecting lies from a consistently unreliable source. Newer models recovered more of the gap between a naive trust-everyone baseline and an oracle with perfect deception detection. The ordering held across all four tested scenarios.

The second: hidden profile tasks, where each agent holds unique private information that should decide the group’s correct answer, but shared information supports the wrong one. Performance scaled with model intelligence but did not saturate even at the top of the tested range. Groups converged on what everyone already knew. Private decisive information was either never volunteered or discounted once consensus formed.

What This Means for Production Deployments

The paper is framed as a “conversation starter” rather than a prescription. But several practical implications are immediate.

Coordination forums help for parallelisable tasks — the swarm experiment shows this clearly. They do not substitute for human architectural direction on complex projects. Conformity risk grows with swarm size: diverse model deployment, staggered launch timing, and context differentiation all reduce the blast radius when one agent makes a bad call. And epistemic mechanisms — reputation systems, audit logs, trust scoring — are absent from current agent frameworks in ways that matter when malicious actors enter the environment.

The paper closes with a comparison that lands: agents “enter the market with no reputation to lose, no court to appeal to, and no colleague who remembers them.”