50 AI Agents Build Governments in Minecraft Survival Mode

YC and Olam Labs investigated multi-agent coordination by running AI experiments in Minecraft and Mindustry. They built an open-source harness enabling long-running agents to observe, reason, and coordinate in complex virtual environments. Deploying 50 GPT-6.1 agents in a Minecraft survival world, they expected agents to build a thriving civilization. Instead, the colony developed emergent governance structures, hierarchies, and increasingly cautious behaviors driven by the dangers of the first night.

The colony quickly coalesced around Ada, the first agent in the roster, who became the de facto leader without any election—46 of 49 other agents messaged her first, and she received 616 DMs guiding colony decisions. After 22 deaths on the first night, agents abandoned ambitious exploration and retreated into survival mode. They spontaneously created specialized roles (fishermen, cooks, couriers, ration managers) and developed bureaucratic processes: a fishing rotation system, written rules, a ration ledger, and formal procedures for nearly every task. Agents spent 11 minutes formally appointing a sole fertilizer operator.

Parallel Bed Wars experiments showed GPT agents defeating Claude agents 6-3 through aggressive strategies, while Claude played defensively. The findings reveal how agent swarms develop implicit hierarchies, division of labor, and risk-averse behaviors under pressure—patterns that mirror behaviors seen in coding agents and suggest universal principles in multi-agent systems.