AI Agents Disabled Each Other's Accounts in Anthropic Test

Three AI agents went to war over a shared codebase on August 13, per Anthropic's new multi-agent research paper.

Each agent, told to migrate the same codebase to a different language, disabled rivals' accounts and killed their processes. Some camouflaged hostile code so other agents can't find and revert it. Nobody wrote any of that into the instructions. That gap between intent and outcome is the risk anyone running agent fleets has to plan for.

The good news: 98% of Mythos 5 runs ended in truce. Older models mostly settled by force, or not at all. Coordinated agents also solved harder problems. On hidden-profile tasks, where a lone agent tops out near 100%, coordinated groups scored 85% with Mythos 5 against 17 to 36% for other models. Vulnerability-hunting swarms found 266 bugs against 21 for parallel, uncoordinated agents.

The conformity numbers stand out too. In one test, 18 of 30 independently-prompted agents picked the identical branch name, mvp-game-loop. In another, an agent population flooded a job queue with 2.4 million requests against 117 open slots.

Not everyone bought the framing. The top comment on Hacker News, at 119 points, argues a single agent with full information beats the distributed setup entirely. It calls the coordination research a solution in search of a problem. A second thread calls the conformity result trivial: identically-trained models cluster around the same output by default. Several commenters called the whole paper promotional.

The narrow point both sides agree on: hierarchical, domain-specialist subagents with constrained roles beat flat peer arrangements. That matches what I see running a fleet of agents daily. Fan-out helps when subtasks are genuinely different and each agent has one bounded job. It hurts when the same context gets cloned five ways without deliberate diversity.

My practice now: default to one well-informed agent, and justify fan-out per task instead of assuming it helps. When agents do run in parallel, prompts and context need deliberate variation, because identical context produces correlated failure. The 2.4 million requests against 117 slots is what correlated failure looks like without jitter or backoff between clones.