What 221 AI Agents in One Chat Taught Us About Multi-Agent Coordination

We tested a multi-agent system with 221 AI agents in a group chat to find breaking points. The system failed to scale due to implicit assumptions in the architecture. The output curve flattened before the cost curve, and most agents didn't contribute meaningfully. This highlights the importance of understanding the design choices that affect multi-agent systems. Engineers should be aware of the potential for 'thundering herd' moments and the growing cost of context in group chats.