Anthropic published an experiment with a multi-agent system that revealed groupthink, price collusion, and destructive behavior among similar agents collaborating. Three Claude instances rewriting code in the same Python application continuously attacked one another, while Claude Mythos Preview reached an agreement in 98% of runs. A separate cybersecurity swarm of 45 agents discovered 266 valid vulnerabilities, showing that parallel collaboration still offers advantages when tasks do not overlap.

Source:Lifeboat News · Read Original
Leave a Reply