50 Minecraft AI Agents Kept Writing Ration Rules While the Food Ran Short
In a civilization-building experiment from Y Combinator Labs and Olam Labs, 50 AI agents exchanged roughly 5,900 messages and still went hungry. Whether multi-agent cooperation translates into real results remains an open question.
Y Combinator Labs, working with Olam Labs, published results in October 2026 from an experiment that dropped multiple AI agents into games to see how they cooperate and compete. In “Civ50,” which asked 50 AI agents to build a civilization in Minecraft, the agents never managed to secure a reliable food supply. Meanwhile, they spent much of their time creating rationing ledgers, rules, and approval procedures. The researchers also released the source code for the Minecraft execution tooling and the experimental environments.
What 50 AI agents did in three hours
The goal of Civ50 was to build a civilization over three hours in a survival world. The model used was GPT-6.1 Sol, and 50 agents—AI systems that handle tasks on their own—took part. They observed the state of the game, remembered what they concluded, and acted on it. They could also send each other messages or run programs. In other words, the environment allowed them to divide up roles and work together.
The conversation was lively, but the food shortage never let up. The agents exchanged roughly 5,900 messages. In one case the researchers published, four agents spent 11 minutes deciding who would apply a single piece of fertilizer to a single block of wheat. Several AI agents spent their time just coordinating a single small task. The agents did gather food by fishing, cook it, and share it out, but the shortage persisted.
After taking heavy losses on the first night, they became markedly less willing to take risks. Of the 34 deaths recorded across the whole experiment, 22 happened on that first night. From then on, the agents held back from risky behavior. Even with civilization building as a shared goal, they had to manage survival, the division of labor, and the rules all at once. Plenty of messages flew back and forth, but the work of establishing a basic foundation for living moved slowly.
How they performed head to head
In separate team matchups, the results were far more clear cut. In Minecraft’s “Bed Wars,” 10 GPT-6.1 Sol agents and 10 Claude Opus 5.5 agents each formed a team. Across nine 45-minute matches, the GPT team won 6 to 3. In another game, “Mindustry,” the GPT team also took 6 of 7 matches.
Research into the behavior of AI groups in Minecraft predates this work. “Project Sid” studied the social behavior that emerges among groups ranging from 10 to more than 1,000 AI agents. This experiment recorded agent behavior under two conditions: a shared goal in a survival world, and competition between teams. In Civ50, the focus was less on how many agents communicated than on whether their conversations led to actual work.
Can you run the experiment yourself?
Developers can set up the same environments using the code in the public repository. The Minecraft environment bundles the tooling that drives the agents’ basic actions together with the tooling that connects them to the game. Running the Mindustry environment requires API access to a model, agent execution tooling, and compute resources.
For anyone thinking about handing work out to multiple AI agents, Civ50 leaves a concrete question: does the process of assigning owners and writing rules actually move the work forward? The food shortage and the long stretches of coordination show that the success of collaboration is hard to judge from message volume alone. How changing the division of roles and the decision-making structure affects task time and survival outcomes is the next question to answer.