Researchers from Y Combinator and Olam Labs deployed large language model-based agents in Minecraft using the Mindcraft framework and Mineflayer scripting tools to study multi-agent dynamics. In a collaborative "Civ50" scenario, 50 GPT-6.1 Sol agents were tasked with building a thriving civilization within three hours. After initial losses to zombies on the first night, agents abandoned ambitious tasks and defaulted to bureaucratic paralysis; notably, four agents spent 11 minutes designating a sole fertilizer operator for a single bone meal on one wheat block. The researchers compared this behavior to patterns observed in coding agents. In competitive Bed Wars matches, 10 GPT-6.1 Sol agents faced 10 Claude Opus 5.5 agents. GPT agents developed aggressive, elaborate strategies and diagnosed a tool bug preventing block-breaking, sharing workarounds, while multiple Claude agents concluded breaking was impossible. In Mindustry, a lightweight RTS tower defense mode, GPT agents won six of seven games. Researchers attributed this to GPTs operating as a cohesive hive mind versus Claude's single-leader, democratic decision-making process. The agent swarms are open-sourced, enabling further experimentation with model mixtures and swarm sizes to investigate coordination, tool use, and emergent social structures in simulated environments.