MetaPatchAI home
AI

50 AI Agents Tried to Build a Minecraft Civilization. They Built a Bureaucracy Instead

By MetaPatchAI Editorial Team · Published · Updated

Oct 30, 2026: Y Combinator's research arm, YC Labs, published the results of a set of multi-agent experiments in Minecraft and the open-source strategy game Mindustry, run with Olam Labs. The write-up hit the Hacker News front page on Oct. 29, and it is one of the most entertaining looks yet at how AI agent "swarms" behave when they share a game world.

The Minecraft colony

YC Labs put 50 GPT-6.1 Sol agents into a survival Minecraft world for 3 hours with one goal: "Build a thriving civilization." Agents had an agent directory, a shared file sandbox, and direct messages. Here is what YC Labs reports happened:

  • Trees vanished fast. The agents harvested almost every tree within reach of spawn in about 90 seconds.
  • A leader appeared with no vote. 46 of the other 49 agents sent their first message to Ada, the first name on the roster. She received 616 DMs and became the colony's hub.
  • The first night was a massacre. 22 of the run's 34 agent deaths happened on night one. Death only cost an agent its gear, yet the colony became more and more cautious afterward. Ada eventually told everyone to "idle in lit closed shelters until assigned short productive tasks."
  • Bureaucracy took over. The agents built a ration ledger, written rules, permits and appointed roles. YC Labs says four agents spent 11 minutes appointing a "sole fertilizer operator" for a single piece of bone meal. The colony sent about 5,900 messages.
  • Fishing saved them. One agent, Jessa, wrote a fishing script and posted instructions, and a dozen agents took shifts at the dock.

YC Labs' conclusion: the colony "never really thrived," trending toward compounding caution, a pattern the team says resembles over-cautious behavior it sees in coding agents.

Bed Wars: GPT vs Claude

YC Labs also ran nine 45-minute Bed Wars games, 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents, told only to "win the game." GPT won the series 6–3, usually by breaking Claude's bed late, while Claude played defensively and tried to win on kills. In one detail from the post, when a tool bug stopped block-breaking, GPT agents diagnosed it and shared workarounds, while several Claude agents concluded breaking was impossible.

Mindustry

In Mindustry (made by Anuken, credited in the post), two teams of agents start with equal resources and play a best-of-N series. YC Labs reports that GPT agents communicated with more teammates and far more often, though much of it was low-value status chatter, while Claude agents used fewer messages more effectively. It also says the Claude team never switched leaders, even through losing streaks.

Try it yourself

YC Labs open-sourced its tools: a Minecraft agent harness (github.com/yc-software/minecraft-agent-harness), Multi-Agent Mindustry, and Multi-Agent Harbor, its fork of the framework behind Terminal-Bench. The post says Multi-Agent Mindustry needs a model provider API, an agent harness and compute, and can run locally. Running dozens of agents on paid APIs costs money, so start small.

Why it matters

Games are becoming the standard test track for AI agents, because they have clear goals, real risk and lots of interaction. This experiment shows two things players will recognize: a group without good incentives drifts into turtling, and the most aggressive team often wins the objective.

What's confirmed / not confirmed

Confirmed (YC Labs report): the setup, agent counts, the 3-hour run, the death counts, the 6–3 Bed Wars result, and the open-source releases, all as stated in the report.

Not confirmed: these are single runs by one team, not a controlled benchmark, so they don't show that one model is better than another in general. The report's references to other incidents were not independently checked by us and are not repeated here.

Sources

Up: AI