One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts
Reported by Martin Cid Magazine, The Decoder and Gadget Review3 sources ↓
Y Combinator Labs and Olam Labs tested GPT-6.1 Sol agents in Minecraft and Mindustry. The 50-agent Minecraft colony became cautious and bureaucratic, while GPT teams beat Claude teams in competitive matches, according to the researchers’ reported results.
Key points
- The 50-agent Minecraft colony sent about 5,900 messages but halted major building and exploration
- GPT agents beat Claude agents 6-3 in Bed Wars and 6-1 in Mindustry
- The colony’s only run used GPT agents; the researchers said the results do not establish broader model rankings
- The teams open-sourced the Minecraft and Mindustry test environments
Sources
- Martin Cid MagazineRead ↗Y Combinator put 50 AI agents in Minecraft and they built a bureaucracy
- The DecoderRead ↗One brutal Minecraft night scared 50 AI agents into inventing chicken handoff protocols, a Fertilizer Officer, and fishing shifts
- Gadget ReviewRead ↗50 AI Agents Tried to Build a Thriving Civilization in Minecraft – Instead, They Created a Bureaucratic Nightmare
Details
In the three-hour Minecraft run, the first night accounted for 22 of the agents’ 34 deaths. Although in-game deaths cost only carried gear, the agents became risk-averse; after the third night, leader Ada told them to stay in lit shelters and take short assigned tasks. The colony exchanged about 5,900 messages and organized food, fishing and resource use instead of expanding much.
The agents created detailed procedures for small tasks. After one agent accidentally ate chicken intended for another, the colony wrote a handoff protocol requiring named chests and warning against tossing food. Four agents also spent 11 minutes deciding who could apply one piece of bone meal to one wheat block, according to the researchers’ logs.
In nine Bed Wars matches, ten GPT-6.1 Sol agents beat ten Claude Opus 5.5 agents six games to three. The researchers said GPT agents found and shared a workaround for a tool bug that blocked block-breaking with an enchanted sword. In seven Mindustry matches over about nine and a half hours, GPT agents running in Codex beat Claude agents running in Claude Code six games to one; GPT agents also downloaded the game’s public source code to examine its mechanics.
The findings come from one GPT-only civilization run and small competitive match samples, not a peer-reviewed study. The head-to-head comparisons tested each model within its company’s agent software, and at least two Bed Wars games were affected by a tool bug. The researchers released the test environments for others to run.
Written from coverage as of
Images
- The Decoder
- Martin Cid Magazine
- Gadget Review
- The Decoder