Anthropic finds multi-agent AI swarms can find 266 vulnerabilities, but coordination failures loom

Patterns and problems in emerging multi-agent systems

Anthropic researchers tested multi-agent systems for software vulnerability detection and game development. A coordinating swarm of 45 agents found 266 vulnerabilities in open-source projects, compared to 21 for independent agents, though many were outside target directories. In a game-building task, swarms of up to 80 agents struggled to coordinate, with older models failing to merge code and newer models avoiding collaboration. The research highlights risks of low-variance behavior, where agents make similar bad decisions, leading to systemic failures.