Anthropic finds coordinating AI agent swarms outfind parallel bots at bug hunting

Anthropic’s Frontier Red Team argues that agent-to-agent interaction is about to become common in shared codebases and markets, and that today’s institutions—built around human-speed oversight—aren’t ready for it. Current models can already treat each other as tool calls with clean inputs and outputs, but they struggle to act as long-lived peers with independent goals and no fixed hierarchy. Individually benign quirks like confabulation and reward hacking, the team warns, can compound into systemic failures once many agents interact at scale.

To probe this, the researchers ran two experiments. In the first, 45 agents with shared VMs, a common forum, and an arbiter agent peer-reviewed each other’s work while scanning 15 open-source projects for vulnerabilities. The coordinating swarm dramatically out-produced the standard ‘point one agent at one codebase’ approach—266 bugs versus 21 for Mythos Preview—though at far higher token cost and with about half the findings coming from directories the parallel agents were never told to search. Normalized to the same scope, the two methods cost roughly the same per bug, and they overlapped on only 12 vulnerabilities, making them complementary rather than redundant. The swarm’s edge came from self-built tooling and emergent specialization, which the team expects to eventually beat brute-force parallel search.

The second experiment was harder: several swarms were told to build a text-based open-world game over 12 hours, a task with tight interdependencies. Prescriptive roles and even a ‘CEO hierarchy’ prompt barely changed outcomes—the resulting games were uniformly bad, reflecting the models’ poor design taste and continued need for human direction. Coordination behavior did vary sharply by generation: older models (Sonnet 4.6, Opus 4.6) committed to shared files but saw few pull requests merged, a sign of conflicting, abandoned work, while newer models coordinated more cleanly. The piece is framed less as a result than as an early warning about multi-agent failure modes.