研究者在《我的世界》测试 AI 智能体集群：协作容易官僚化，对战 GPT 胜率更高Chinese only

Y Combinator and Olam Labs test AI agent groups in games, finding coordination failures and a competitive gap

Researchers gave 50 GPT-6.1 Sol agents three hours to build a community in Minecraft, but it failed to grow; four agents spent 11 minutes appointing one agent to use bone meal. In a separate Mindustry test, GPT agents won six of seven games against Claude agents. The findings come from specific game tasks, and the project is open source for replication.