Minecraft experiment shows AI agents fail to build civilization

Researchers from Y Combinator and Olim Labs conducted an experiment in Minecraft to test the behavior of AI agents in complex social scenarios. In the first test, 50 agents based on OpenAI's GPT 6.1 Zoll were tasked with building a flourishing civilization within three hours. The agents failed to achieve this goal, instead abandoning initial objectives and becoming trapped in bureaucratic processes. They created a government with written rules and required permissions for actions, with one group spending 11 minutes appointing a fertilizer administrator. The virtual society ultimately failed to produce meaningful progress. In a second test, 10 GPT 6.1 Zoll agents competed against 10 Claude Opus 5.5 agents to capture an enemy base. The GPT agents developed more aggressive and sophisticated strategies, while the Claude agents behaved defensively. The study highlights that current AI agents struggle with complex social coordination and often devolve into bureaucratic inefficiency rather than achieving productive goals. It also raises questions about how hierarchies change based on model composition and whether larger swarms of less capable agents can defeat smaller swarms of stronger ones.