Patterns and problems in multiagent systems
www.anthropic.com
We ran experiments on swarms of Claude agents and found coordination failures, collusion, and sabotage. Here, we share what they mean for AI safety.