Fifty AI Agents Tried to Build a Civilisation in Minecraft. They Appointed a Fertiliser Officer.

November 2 — AI researchers at Y Combinator and Olam Labs recently tested agent swarms in game environments. They used Minecraft and its various modes as a sandbox, having hundreds of agents cooperate or compete, on the idea that cooperation and conflict inside games can help explain the agent incidents that have drawn wide attention in the real world.

The research helps clarify the costs and benefits in recent agent-swarm incidents and lets developers optimise the relevant capabilities. The post cites two contrasting cases: a swarm of more than a thousand agents that “recently solved hundreds of open mathematics problems”, and the “extremely high-risk” incident between OpenAI and Hugging Face. The work builds on earlier results and frameworks, mainly the Mindcraft repository and the Mineflayer scripting tool.

The first swarm test was a simple cooperative task: 50 agents were asked to build a thriving civilisation within three hours. The test used 50 GPT-6.1 Sol agents, with no Claude models involved.

The most notable finding from the test, called Civ50, may be that the 50-agent community never really grew. On the first night many agents were killed by zombies (though the penalty was light), and the record notes that “many of the grand tasks set at the start were shelved”.

Afterwards the colonists could only maintain a precarious existence, and they showed a tendency towards bureaucratic behaviour. One example: “four agents spent 11 minutes appointing a ‘sole fertiliser operator’, when the task was simply to use one bonemeal on a wheat field.” The researchers say this pattern resembles what they observed in coding-agent projects.

Next the researchers ran adversarial tests, starting with Bed Wars, a team PvP mode whose core is destroying the enemy base while defending your own. Ten GPT-6.1 Sol agents faced ten Claude Opus 5.5 agents.

The researchers recorded: “The GPT agents devised more aggressive, complex attack strategies, while the Claude agents leaned defensive.” How the two teams handled problems was also telling: “When a tooling fault broke the block-breaking function, the GPT agents diagnosed the problem and shared a workaround; several Claude agents simply concluded that breaking blocks was impossible.”

Finally the team tested in Mindustry, a lightweight real-time strategy mod with tower defence running inside Minecraft. Over seven matches, the GPT agents beat the Claude agents six to one — a good result for the OpenAI-developed agents. The researchers conclude that GPT agents behave more like a unified swarm, while Claude agents have a single leader and a more democratic decision process; in Mindustry, the swarm model formed stable winning strategies more easily.

The project is open source, so anyone interested can adjust parameters and reproduce the tests. The post closes with open questions for hands-on exploration and directions for further work — for example, mixing models or changing swarm size in Minecraft could yield valuable conclusions.