Y Combinator Tests AI Agent Swarms in Minecraft, GPT Beats Claude 6-1

Researchers at Y Combinator and Olam Labs tested AI agent swarms inside Minecraft, letting over a hundred agents team up or fight each other. The goal is to understand how real-world agent swarm events unfold. The project is open source.

In the Civ50 test, 50 GPT-6.1 Sol agents had to build a thriving civilization within three hours. Many got killed by zombies on the first night, big plans were shelved, and they even became bureaucratic—four agents spent 11 minutes appointing a single 'fertilizer operator' for one wheat field.

In head-to-head battles, GPT-6.1 Sol faced Claude Opus 5.5: GPT played aggressive offense and shared fixes when tools broke, while Claude stayed defensive and often ruled out broken tasks. GPT won six of seven tower defense rounds. Researchers found GPT acts like one unified swarm, while Claude has a single leader.