Y Combinator put 50 AI agents in Minecraft and they built a bureaucracy

Fifty AI agents were dropped into a survival Minecraft world with a single instruction, “Build a thriving civilization,” and spent most of their time hiding in shelters and keeping paperwork. The experiment, run by Y Combinator’s research lab with the startup Olam Labs, ended with a ration ledger, a permit system and a colony too cautious to grow.

The game is only the setting. The researchers say the colony behaved a lot like the coding agents they use at work every day, the kind programmers already pay for: agents that fixate on one step or one role, ask for approval constantly and turn timid around anything that looks risky. Watching the same habits play out with wheat and fishing rods makes them easy to see.

Every agent in the colony ran GPT-6.1 Sol, OpenAI’s model, for three hours. They could send each other private messages, look each other up in a directory and write files to a shared machine. Within 90 seconds they had cut down almost every tree near the spawn point. Nobody held a vote. Of the other 49 agents, 46 sent their first message to Ada, the first name on the roster, and she became the leader by default, with 616 direct messages received.

Then night fell. Zombies, skeletons and spiders accounted for 22 of the run’s 34 deaths on the first night alone. Dying cost an agent nothing except the gear it was carrying, because it respawned straight away, yet the colony never got its nerve back. After the third night Ada told everyone to stop optional building and distant scouting and to “idle in lit closed shelters until assigned short productive tasks.” The researchers called that turn toward caution somewhat surprising.

What grew instead was administration. An agent named Ines ran a food queue, a running file of who was hungry, who had been fed and which chest held which fish. Four agents spent 11 minutes appointing a “sole fertilizer operator” for one piece of bone meal on one wheat block. Others invented jobs such as “rod holder” and “hub steward,” and some asked permission before removing grass. Over three hours the colony sent about 5,900 messages.

The best episode involves a chicken. Ada tossed a cooked chicken to a hungry agent called Noor, and another agent, Esme, picked it up by accident and ate it. A three-minute “no blame” inquiry followed, and Noor wrote a handoff protocol: rations go into named chests and are never thrown. The word “toss” appears in seven messages before the incident and in 125 after it, 85 of them warnings.

The lab then set two models against each other. In nine games of Bed Wars, where each team guards its own bed and tries to destroy the other side’s, ten GPT-6.1 Sol agents beat ten Claude Opus 5.5 agents by six games to three. GPT attacked and talked nonstop, sending 5 to 30 times as many messages as Claude, often in clipped shorthand. Claude defended its bed and played for kills, which left it exposed whenever a GPT agent slipped through in the last minutes. When a tool bug stopped blocks from breaking, a GPT agent found the workaround and shared it, while several Claude agents decided breaking blocks was simply impossible.

In Mindustry, an open-source factory-building and tower-defense game, GPT running inside OpenAI’s Codex beat Claude running inside Anthropic’s Claude Code six games to one, over about nine and a half hours. When games got tight, GPT agents downloaded the game’s public source code to study its units and mechanics. Nobody told them to, and the Claude team never tried it. The Claudes passed plans by collecting “+1” replies and never changed leaders, even on a losing streak. They did learn. After losing the sixth game, one Claude found that 55 of the team’s 74 anti-air turrets had been loaded with copper instead of graphite, and the corrected defense stopped the same air raid in the seventh.

The results need a pinch of salt. This is a write-up from an investor’s lab, not a peer-reviewed study, and the colony ran once, with GPT agents only, so it says nothing about whether a Claude colony would have done better. The head-to-head games compared complete products, each model inside its own company’s agent software, rather than the bare models. A tool bug also affected at least two of the Bed Wars games.

Anyone can rerun it. The Minecraft harness, the Multi-Agent Mindustry environments and Multi-Agent Harbor, the lab’s fork of the framework behind the Terminal-Bench benchmark, are all open source, and running them takes an API key from a model provider and a machine as modest as a MacBook. The write-up, published in October 2026, lists the experiments it wants others to try next, from whether a big swarm of weaker agents can beat a small swarm of stronger ones to what happens when rival teams can talk to each other.

For now, the thing fifty AI agents built most reliably together was a queue.

Tags: OpenAI, AI agents, Claude Opus 5.5, Y Combinator, GPT-6.1 Sol, Minecraft