50 AI Agents in a Game Wrote Their Own Rules for Sharing Food
Y Combinator Labs and Olam Labs ran experiments on AI cooperation and competition in Minecraft and Mindustry. They also released the harness and the test environment, so anyone can run the same experiments.
In October 2026, Y Combinator Labs and Olam Labs released the results of experiments in which large numbers of AI agents were set loose simultaneously inside Minecraft and Mindustry. In the experiments, AI agents that handle tasks on their own either cooperated to survive or competed in teams. Both organizations also published the Minecraft harness and the Mindustry environment to public repositories.
How did 50 AI agents cooperate?
In the Minecraft survival experiment, 50 GPT-6.1 Sol agents created their own rules for distributing food and divided up roles among themselves. Over three hours, they exchanged roughly 5,900 messages. Beyond messaging, the agents also made use of a shared workspace. To move around inside the game, they wrote and executed the commands they needed on their own.
Deaths clustered early. Of 34 total deaths, 22 occurred on the first night. After that, the agents voluntarily cut back on risky activity. What the researchers were watching was this process of change more than gameplay skill: the flow of multiple AI systems exchanging information in a shared space, setting rules, and adjusting how they behave.
Minecraft has been used in AI research before. What sets this release apart is its focus on cooperation and competition among multiple agents. In Mindustry, each agent was given its own separate runtime environment and the same set of game controls. The published Mindustry environment supports team-versus-team matches as well as survival runs with up to 20 agents at once.
Who won the team matches?
In Minecraft Bed Wars matches that pitted model against model, the GPT-6.1 Sol team came out ahead, winning 6 of 9 games against the Claude Opus 5.5 team. In Mindustry, the GPT-6.1 Sol team won 6 of 7 games. Running the full Mindustry series took about 9 hours and 30 minutes.
What do you need to run it yourself?
The first step is wiring the game environment to an AI model. The Minecraft harness requires an API key from a model provider, a Minecraft server, and runtime software at or above a specified version. The Mindustry environment likewise needs a way to reach a model, an agent runtime, and the compute resources to run it.
The approach is to pull the tools from the public repositories and stand up the experiment environment yourself. Users are responsible for connecting the model and operating the server. The announcement did not state the cost of a single run. Model usage fees and compute costs are items to weigh before you start.
The significance of this release lies in shipping both the tooling and the environment needed to look inside the process of multiple AI systems cooperating. The focus is also shifting, from how well a single AI plays a game to how multiple AI systems divide up the work. The things to watch going forward: whether the same role division and risk avoidance show up under different conditions, and how much time and money it takes to reproduce them.