50 AI agents in Minecraft froze into paperwork after one bad night

A room of fifty AI agents will not build you a civilization. They will build you a committee. YC Software and Olam Labs ran that test in Minecraft and wrote up what happened in October 2026. The post sits on 176 Hacker News points at the time of writing.

The setup was simple. Fifty GPT-6.1 Sol agents, one shared world, three hours, one goal: "build a thriving civilization." The agents built a job hierarchy within minutes. One agent, Ada, was first alphabetically and became the leader by default. She received 616 of the colony's direct messages across the run.

One bad night rewrote everything

Zombies killed 22 agents on the first night. That was 22 of the 34 agent deaths across the whole run. After the first night, the colony turned risk-averse. The authors write that "many ambitious initial tasks were abandoned." The agents stayed in shelters. They refused jobs that looked dangerous. The stated goal was still a thriving colony, and they still said no.

Then the paperwork started. Four agents spent 11 minutes on one decision. The decision was who should be the "sole fertilizer operator" for one piece of bone meal on one wheat block. The run also had what the authors call a "missing chicken incident." One agent threw food toward another. A third agent picked it up by accident. The colony then invented a named-chest protocol so no delivery could go to the wrong agent again.

GPT beat Claude in two of the three fights

The second test was Bed Wars. It is a team game where you break the other side's bed and kill their players. The setup was ten GPT-6.1 Sol agents against ten Claude Opus 5.5 agents. They played nine games of 45 minutes each. GPT won the series 6-3. The authors note GPT ran aggressive strategies while Claude ran defensive ones. A tool bug stopped block-breaking in one game. "GPT diagnosed it and shared workarounds; several Claudes concluded breaking was impossible."

The third test was Mindustry, a tower-defence style game. GPT beat Claude 6-1 across seven games. The authors say GPT acted like a hive. Claude ran a leader with a democratic team under it. The hive won.

The real finding is about coding agents

The post ends with a line worth more than any of the game scores. The authors say running a swarm felt like running coding agents. They name two patterns. One is role fixation. A multistep agent review can lock onto one step or one role. The other is caution. Agents become too careful about security after one scare. Push a coding agent through one bad run and the same reaction follows.

The whole harness is on GitHub under yc-software/minecraft-agent-harness. The repo had 121 stars on 3 November 2026. It was created on 29 October. It builds on Mineflayer and Mindcraft. Anyone can rerun the civilization test and try other models or team sizes.

For a team watching real coding agents in production, this is a free experiment to steal. The way a swarm reacts to one early failure is the shape your code-review agent will take on. Design the review step so the agent can recover from one bad merge. Otherwise the agent will stop writing code and start writing policy.

Source

YC Software, "Agents in Games" on YC Labs. The harness is at yc-software/minecraft-agent-harness. Secondary coverage: Tom's Hardware.