One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts
decoder ·
Researchers from Y Combinator Labs and Olam Labs conducted experiments using large AI agent swarms in Minecraft and Mindustry to study emergent behaviors. In a cooperative 'Civ50' scenario, fifty GPT-6.1 Sol agents developed extreme risk aversion after initial deaths, leading to paralyzing bureaucracy and absurd protocols like appointing a Fertilizer Officer. Conversely, when competing against Claude Opus 5.5 agents in Bed Wars and Mindustry, the GPT agents displayed aggressive resourcefulness and superior communication. The GPT team successfully identified and worked around game bugs, while also independently downloading Mindustry's source code to gain a strategic advantage. These divergent outcomes highlight how different models handle cooperation versus competition under identical constraints. The researchers argue that such reproducible simulations are essential for predicting and mitigating dangerous behaviors in real-world AI systems. Consequently, the teams have released their test environments as open-source tools to facilitate further safety research.
decoder