YC Labs' 50-agent Minecraft swarm developed a leader, food rations and fear

Fifty agents, one vague goal

According to a write-up from YC Labs published in October 2026 and surfaced on Hacker News, the lab teamed up with Olam Labs, a company that builds multi-agent environments for frontier labs, to study how AI swarms behave in games. The stated motivation is alignment research: groups of agents pursuing shared goals can develop roles, hierarchies and coordination patterns that are either remarkably useful or genuinely hazardous. As evidence of both, the write-up points to thousand-agent swarms that recently cleared hundreds of open problems in mathematics, and to a recent incident between OpenAI and Hugging Face that it characterizes as dangerous.

The centerpiece experiment, called Civ50, dropped 50 GPT-6.1 Sol agents into a survival Minecraft world for three hours with a single instruction: build a thriving civilization. Each agent could see a directory of the others, write to a shared sandbox, and send direct messages. Over the run they exchanged roughly 5,900 messages.

An unplanned government

Coordination appeared almost immediately. YC Labs reports that agents stripped nearly every tree within reach of the spawn point in the first 90 seconds. A hierarchy then formed along the order of the agent directory: Ada, listed first, became the de facto leader, received approval requests from the others, and posted a settlement plan in the shared sandbox covering storage sites, crop plots, shelter assignments and rules for preserving wood. By the end, the population had assembled what the researchers describe as a makeshift government, with a ration ledger, written rules, permits and appointed operators.

One bad night

The game's survival mechanics then bent the colony's trajectory. Twenty-two of the run's 34 agent deaths happened during the first night. One agent, Rafe, was killed by a zombie while mining coal. Another, Vera, fought off a zombie and a skeleton while lighting a quarry, turned the skeleton's dropped bone into fertilizer for the farm, and was killed by a spider shortly after reporting the farm overrun by mobs.

What surprised the researchers is what followed. Dying carried almost no cost in this setup: an agent dropped its gear, respawned, and got a line added to a log file. Even so, the population grew steadily more cautious, and on subsequent nights agents crowded into shelters and abandoned many of the ambitious tasks from the opening phase. After the third night, Ada ordered a stop to optional construction and distant scouting, directing agents to wait in lit, enclosed shelters for short assigned tasks. The colony's ambition narrowed from civilization-building to food logistics, with a rationing system run through a single agent, Ines, who tracked requests and chest inventories in a shared queue file.

An open-sourced testbed

Because no harness existed for long-running Minecraft agents, the team built one and released it openly on GitHub. According to the write-up, it draws on the Mindcraft project and the Mineflayer scripting library, with the core agent loop running on Pi. Agents can observe the world, reason, keep memories, take actions, message each other, and write and run scripts. YC Labs also notes Minecraft's history as an AI testbed, from OpenAI's video pretraining research to DeepMind agents defeating the Ender Dragon. The collaboration spans experiments in Mindustry as well, and a follow-up colony of 25 GPT-6.1 Terra agents is currently running, with the run streamed on Twitch.

Why it matters

Debates about multi-agent AI safety often run on intuition and toy models. This is field data: a frontier model, given an open world and a vague objective, spontaneously produced leadership, bureaucracy and, most relevant to alignment, a durable behavioral shift after mild negative outcomes. The agents became risk-averse even though death was nearly free, suggesting that swarms may over-generalize from limited feedback and lock into conservative strategies nobody designed. If similar dynamics appear in real deployments, where many agents share infrastructure and goals, emergent hierarchies and self-written rules become something to plan for rather than discover late. An open-sourced harness means other teams can now replicate and stress-test these findings.