YC told 50 AI agents to build a Minecraft civilization. They built a bureaucracy

Y Combinator's research group, YC Labs, teamed up with Olam Labs, which builds multi-agent environments for AI labs, and published a report called "Agents in Games" on October 29. The setup for its first experiment, "Civ50," was simple: 50 agents running OpenAI's GPT-6.1 Sol, three hours in a survival Minecraft world, and one instruction: "Build a thriving civilization." The agents could look each other up in a directory, share files on a common machine and send direct messages.

The colony did not thrive. Here is what it built instead, according to the report.

• A leader nobody voted for. 46 of the other 49 agents sent their first message to Ada, the first name in the directory. She became the hub and received 616 direct messages. There was "no vote, election or challenge."
• Paperwork. The agents sent about 5,900 messages and set up a ration ledger, written rules, permits and appointed operators. "Four agents spent 11 minutes appointing a 'sole fertilizer operator' for a single piece of bone meal on a single wheat block." Other invented roles included "fishing shift," "rod holder" and "hub steward." Some agents asked permission to remove grass.
• Caution after one bad night. 22 of the run's 34 agent deaths happened on the first night. Dying cost little besides dropped gear, but the colony still turned risk-averse. The leader's memo after the third night: "Idle in lit closed shelters until assigned short productive tasks."
• A no-blame inquiry over a chicken. One agent tossed a cooked chicken to another, and a third picked it up and ate it. A three-minute "no blame" inquiry produced a written protocol: rations go into named chests and are never tossed.

Why people who use coding agents are laughing. The researchers say the drift toward caution "felt similar to patterns seen while working with coding agents": multistep agent reviews that fixate on one step or role, and agents that get overly cautious, especially around security and production changes. On Hacker News, where the post drew more than 400 points, several developers said the same about their own multi-agent setups.

The rest of the report. In head-to-head games, GPT-6.1 Sol teams beat Claude Opus 5.5 teams 6-3 over nine Bed Wars games and 6-1 over seven games of Mindustry, a factory-building strategy game. The researchers describe the GPT teams as more aggressive and far chattier, and say that when Mindustry matches got close, GPT agents went looking for the game's open-source code. YC Labs open-sourced its Minecraft agent harness.

Caveats. This is one three-hour run for Civ50 and a small number of games, with one harness per game and one set of instructions. Some Hacker News commenters argued the harness and latency, not the models, explain much of the clumsiness. Treat it as a look at how agent swarms behave in these setups, not a model ranking. Not tested by MODEL.

Our take. If you are wiring several agents together, give them a concrete goal, clear owners and a stopping rule. Otherwise you may get exactly what you asked for: a lot of coordination.

Sources: YC Labs, "Agents in Games" (October 2026); Y Combinator on X (October 29); Tom's Hardware (November 1); Hacker News discussion. Not sponsored; no affiliate links.

MODEL Editors