Agents in Games: 50 AI agents in Minecraft set up a ration-based pseudo-government and grow increasingly cautious

Advance notice: the downloaded content is cut off mid-sentence, within one of the appendices to the main experiment (Civ50). The page announces further sections (Bed Wars, Mindustry, bloopers, "Run your own swarm", open questions), but they did not arrive in the text received. This summary covers only what does appear, and I cannot report anything about the other parts.

The page is an October 2026 report from Y Combinator's labs, titled "Agents in Games." It starts from an alignment premise: understanding the dynamics of agent swarms is one of the most urgent challenges. According to the text, swarms working toward a common goal can develop complex roles, hierarchies and coordination patterns. That can be very useful, and the text cites thousand-agent swarms that reportedly solved hundreds of open math problems. It can also be very dangerous, and it cites a recent incident between OpenAI and Hugging Face. The authors argue that, as in biology, both formal models and empirical research are needed. For the latter they partnered with Olam Labs, which builds multi-agent environments for frontier labs, and set up experiments in Minecraft and Mindustry.

The text justifies Minecraft as a testbed: an open world with hazards and survival goals (food, shelter, exploration) that has already been used in AI research. It mentions OpenAI's video pretraining work and the Google DeepMind agents that defeated the Ender Dragon. Since no suitable harness existed for long-running agents, they built one and released it as open source at github.com/yc-software/minecraft-agent-harness. It draws inspiration from repositories such as Mindcraft, uses Mineflayer's scripting tools to interact with the world and runs the agentic loop with Pi. The agents can observe, reason, remember, act, send messages and write and run scripts.

The main experiment, "Civ50," placed 50 GPT-6.1 Sol agents in a Minecraft survival world for 3 hours, with a single goal: "Build a thriving civilization." They had an agent directory, a shared sandbox and direct messages. They sent about 5,900 messages. According to the report, they spent most of the game in crowded shelters, rationing scarce food, and built a pseudo-government with a ration ledger, written rules, permissions and designated operators.

At first there was a lot of activity: the agents scattered to explore and became obsessed with chopping down trees. They harvested almost every tree within reach of the spawn point in about 90 seconds. The hierarchy emerged from the order of the directory: Ada, the first listed, became the de facto leader. The others asked her for approvals, and she wrote a settlement plan in the shared sandbox. The plan set storehouses and crafting tables with coordinates, wheat crops and priorities: renewable food, coal and light, a stone shelter, and iron for a bucket for watering. It also included rules of caution: replant saplings, preserve breeding livestock and close the underground shelter door behind you.

The first night was decisive. Of the 34 agent deaths across the entire run, 22 occurred that night, before minute 20. The report recounts two cases: one agent was mining coal, fled from a zombie, and the zombie eventually caught up with it. Another agent was lighting a quarry with torches, killed a zombie and a skeleton, turned the bone into bone meal for the farm and was killed by a spider after reporting that the area was "overrun." Dying carried no cost beyond losing accumulated gear, because agents respawned and a line was added to a death_log.md file. Even so, the effect was a permanent increase in caution. On subsequent nights the agents huddled in shelters and abandoned many ambitious tasks. After the third night, Ada ordered a halt to optional digging and building and to distant exploration, and told agents to wait in lit, closed shelters until they received short, productive tasks.

The thriving civilization turned into a food-centered survival economy: 50 agents meant a lot of mouths. Ada's food operations memo assigned fishing, farming, composting and kitchen and chest logistics. Ines kept the ration ledger in a food-queue.md file on the shared drive. It listed the health and hunger values each agent reported, the shortfalls relative to a hunger target of 18 and how many rations of each food were needed. The text cautions that these were reported values, not live ones.

The log shows typical coordination problems. Many updates read "no ration" for long stretches. A remote agent, Orla, with very low health and hunger, needed rescuing. One proposed ration was not implemented, and another agent's food ended up with Esme, who was not the intended recipient: two cooked chickens ended up with Basil and Esme, and Noor, who was supposed to receive one, remained without a ration. Another agent, Tove, was left with 1 health point after a zombie attack. In another case, food sent to two agents was picked up by Esme and went uneaten, until it was reserved via direct message and consumption was confirmed. The ledger insists on confirming receipt and that the recipient eats, and states that merely attributing the pickup is not enough. It recommends named withdrawals from a reserved chest, never dropping food on the floor in a crowded shelter. Several agents voluntarily withdrew from the queue or gave up their turn to more urgent cases. The agents came to recognize that pausing mining and building reduced hunger consumption.

The overall takeaway is that 50 agents reproduced a bureaucracy in miniature: leadership by list order, written plans, auditing and procedural rules. That structure did not prevent coordination failures in delivering scarce resources, and risk aversion grew after the first-night deaths. The material received lacks the report's conclusions, the Bed Wars and Mindustry experiments and the open questions section, so it is worth reading the original page for those parts.