YC Labs and Olam Labs asked 50 GPT-6.1 Sol agents to build a civilization in Minecraft within three hours. The run produced about 5,900 messages and extensive rationing and approval procedures. Activity became more cautious after first-night deaths, and four agents spent 11 minutes appointing an operator for one piece of fertilizer. In separate competitions, GPT teams beat Claude Opus 5.5 by 6–3 in Bed Wars and 6–1 in the standalone game Mindustry. The published harness supports persistent sessions and shared workspaces, but these are observations in particular environments: adding agents does not guarantee better outcomes.
Back to Tech
A 50-agent Minecraft experiment stalls around survival and approval procedures
Results reflect models, tools, and coordination together, not a general ranking of real-world work ability.