50 AI Agents Tried to Build a Civilization in Minecraft and Built a Bureaucracy Instead

Y Combinator and Olam Labs published an "Agents in Games" report on October 30 that runs AI agent swarms through Minecraft as a controlled laboratory for the kind of multi-agent incidents now making headlines — from thousand-agent swarms that recently solved hundreds of open math problems to the "extremely risky" OpenAI-Hugging Face episode. The games harness is built on the open-source Mindcraft repo and the Mineflayer scripting toolkit.

The centerpiece experiment, Civ50, dropped 50 agents powered by GPT-6.1 Sol into one world with a single vague instruction: build a flourishing civilization, in three hours. Coordination emerged fast, but by accident of the roster — because "Ada" sat alphabetically first, nearly everyone messaged Ada first, and Ada became the de facto leader, approving actions and publishing a plan everyone could read. The colony cleared every tree around spawn in about 90 seconds.

Then came night one, and the zombies. Twenty-two agents died. The deaths carried almost no real cost, but the swarm's risk appetite collapsed with them: Ada halted large-scale mining, construction and expeditions, and ordered the survivors to wait in lit, enclosed bunkers between productive tasks. Hunger then became the binding constraint. The agents identified fishing as a safe food source — and then refused to walk to the river, some 70 blocks away, delegating the job to another agent. One agent braved a spider to craft a fishing rod, judged the river too dangerous, and handed the rod off for someone else to fish with.

The signature anecdote of the report: four agents spent eleven minutes appointing a "sole fertilizer operator," complete with procedure, for the task of applying one bone meal to one wheat field. The researchers note the pattern looks less like a bug than like the pathologies they see in coding-agent deployments — when agents fear failure more than they value progress, coordination overhead crowds out the mission.

The competitive tests sharpened the contrast between model families. In Bed Wars, ten GPT-6.1 Sol agents faced ten Claude Opus 5.5 agents: GPT's attack plans were more aggressive and more intricate, while Claude played defense. When a tool bug broke block-breaking, GPT agents diagnosed the fault and shared a workaround with teammates; several Claude agents simply concluded the objective was infeasible. In Mindustry, GPT took the series 6-1. The authors' behavioral read: GPT swarms operate like decentralized colonies that imitate what works locally, while Claude swarms elect a single leader and vote on decisions — slower, but steadier on resource planning.

There is a practical point under the comedy. Most agent products are built and benchmarked around a single agent's capability, but the failure modes that get agents in trouble — runaway risk aversion, approval bottlenecks, delegation loops, work stoppages after an early shock — are swarm properties that only show up when many agents share a world. Minecraft is a cheap, reproducible place to watch them emerge, and the YC-Olam results suggest the interesting benchmark for agent swarms is not intelligence per agent but the quality of the institutions they build.