YC Labs ran AI agent swarms in Minecraft

YC Labs, partnering with multi-agent environment builder Olam Labs, ran a series of experiments placing AI agent swarms inside Minecraft and Mindustry to study multi-agent coordination in the wild. The first major experiment, 'Civ50,' dropped 50 GPT-6.1 Sol agents into a survival world with a single directive: build a thriving civilization. Within 90 seconds, agents had stripped nearly every accessible tree near spawn. A de facto leader, 'Ada,' emerged purely based on directory order — other agents messaged her for approvals, and she authored a shared colony plan covering food storage coordinates, farming rules, wood conservation, and shelter logistics. The colony developed a ration ledger, written rules, and appointed operators — a spontaneous pseudo-government. Roughly 5,900 messages were exchanged over the three-hour run. Because no suitable agent harness existed for long-running Minecraft play, the team built and open-sourced one, built on Mineflayer scripting tools with an agentic loop powered by Pi. The broader motivation is AI alignment: understanding how agent swarms develop roles, hierarchies, and emergent coordination — capabilities the post notes can be both highly useful (thousand-agent swarms solving open math problems) and potentially dangerous (citing an unspecified OpenAI–Hugging Face incident).