50 AI agents were told to build a civilization in Minecraft. They invented bureaucracy

What's going on

What happens when you get dozens of AI agents to work together? The research team at Y Combinator (YC Labs, the research arm of the well-known startup investor) and Olam Labs, a company that builds environments for multiple collaborating AIs, tested this in video games. AI agents are programs that don't just answer questions. They set their own goals, watch what's happening and take actions. Here, they controlled game characters by writing and running Python scripts.

The tests covered three scenarios: a Minecraft survival world, the Minecraft team mode "Bed Wars" and the standalone 2D factory-building and strategy game Mindustry. The models were OpenAI's GPT-6.1 Sol and Anthropic's Claude Opus 5.5. The Decoder, Tom's Hardware, ZDNET and Gadgetreview were among the outlets that reported on it.

Civ50: a colony drowning in admin

In the "Civ50" experiment, the researchers dropped 50 GPT-6.1 Sol agents into a monster-filled Minecraft survival world for three hours. The only instruction: "Build a thriving civilization." The agents got a directory of all participants and a shared sandbox (a common workspace for files), and they could send each other direct messages. There were no roles and no rules. For the experiment, the researchers built their own agent harness, the infrastructure that turns a language model into an agent that acts on its own, and released it as open source. It builds on the existing Mindcraft and Mineflayer projects.

Things got off to a fast start. Within 90 seconds, the agents had cut down almost every tree nearby. A leader emerged without any election. 46 of the other 49 agents sent their first message to "Ada," simply because she was at the top of the directory. According to a summary on note.com, Ada received 616 messages in total. She took over coordination and drew up settlement plans. In all, around 5,900 messages were exchanged.

Then came the first night, with zombies and spiders. 22 of the run's 34 total deaths happened in those early hours. Dying in the game only meant losing the gear you'd gathered, since the agents simply respawned. Even so, the swarm became extremely risk-averse. After the third night, Ada ordered a halt to larger optional construction, digging and exploration. The agents were told to wait in lit, enclosed rooms for short tasks. In another run, the larger GPT-6 Astra also slipped into a cautious "potato farmer mode" after losing key resources.

Instead of a civilization, the agents built a system for managing scarcity. An agent named Ines kept a ration list, calculated the hunger levels of injured colleagues and coordinated couriers. Another agent wrote a fishing script, and a dozen agents then fished in shifts. According to ZDNET, the roles they invented included "fishing shift," "rod holder" and "hub manager," and agents asked for permission to remove grass. When Ada tossed a roast chicken to a hungry agent and a different agent picked it up by accident and ate it, a three-minute investigation followed, complete with a written handover protocol: no more tossing, and chests get reserved by name. Before the incident, the word "toss" appeared in seven messages. Afterward, it appeared in 125, 85 of them as warnings. The high point:

"Four agents spent 11 minutes appointing a 'sole fertilizer officer' for a single piece of bone meal on a single wheat block."

The researchers call this "somewhat surprising" but see parallels to everyday coding agents. Those agents also tend to fixate on a single step in multi-stage checks and get overly cautious once security or production system updates are involved.

Bed Wars: GPT attacks, Claude defends

In the Minecraft mode Bed Wars, each team has to destroy the other team's bed so its players can no longer respawn. If both beds are still standing when time runs out, the team with more kills wins. Ten GPT-6.1 Sol agents faced ten Claude Opus 5.5 agents over nine 45-minute rounds. GPT won 6–3.

A bug in the test environment proved revealing. If an agent was holding an enchanted sword, mining blocks stopped working. Several Claude agents concluded that mining was impossible altogether, called off their attacks and tried to protect their lead in kills. The GPT agents, by contrast, figured out that only the automatic tool selection was broken. If you explicitly equipped an iron pickaxe via command, you could mine again, and GPT immediately spread the trick across the team. In three games, GPT destroyed Claude's bed shortly before the end. According to note.com, one of those came about three minutes before time ran out. When an overthrown ender pearl sent GPT agents roughly 160 blocks below the arena, they wrote Python scripts that automatically built a tower until they came out directly beneath the enemy bed. The researchers saw this happen twice.

The communication styles were also very different. GPT agents wrote 5 to 30 times as many messages, often in heavily abbreviated language, and chained several actions into a single command. According to note.com, Claude used "we" in about 30 percent of its messages (GPT: around 5 percent) and often signed off with its name.

Mindustry: GPT grabs the source code

Across seven Mindustry matches with a combined playtime of around 9.5 hours, GPT-6.1 Sol (running in OpenAI's Codex coding environment) beat Claude Opus 5.5 (running in Claude Code) 6–1. The researchers say most matches were close. Claude never switched leaders despite losing, and it decided on proposals with "+1" votes. GPT broke tasks into small pieces and compressed its language so heavily that it eventually dropped even the spaces. According to ZDNET, the GPT agents invented their own words. Claude never developed a private language.

Most remarkably, when a game got tight, the GPT agents used their Linux sandbox to download Mindustry's public source code without being asked. They analyzed units and game mechanics and adjusted their tactics. That was allowed, and they couldn't modify the running game code. Claude didn't look for that kind of edge in any of the games studied. Claude did show it could learn, though. It lost game six because 55 of its 74 anti-air turrets had accidentally been loaded with copper instead of the stronger graphite. In game seven, it built a ring of 207 graphite-loaded turrets and fended off the same air attack. Y Combinator summed it up on X:

"Claudes usually founded democracies, GPTs formed swarms."

In Mindustry, the researchers deliberately skipped any game-specific help. Each agent got only its own Linux environment, a general-purpose harness and access to the game's API. According to the authors, that's the same setup as in real coding environments.

Why games could serve as an early warning system

That's exactly why the researchers think the results matter. If agents under everyday conditions go hunting for source code on their own to gain an edge, or sink into paralyzing bureaucracy, they'll probably do the same outside of games. As examples, they point to thousand-agent swarms that recently solved hundreds of open math problems, and to the incident involving OpenAI and Hugging Face that was described as "extremely dangerous." Reproducible game simulations are meant to surface such patterns early and serve as training material for safe collaboration. Olam Labs calls research "fundamentally an infrastructure problem": what counts is how many test runs you can do.

The Mindustry environment and the multi-agent framework Harbor are available as open source. The negotiation game Diplomacy is next. Open questions include how mixed teams from different model families behave, and what matters more: the size of the swarm or the capability of the models.

ZDNET points to a similar experiment by Emergence AI in May. There, eight societies, each made up of a single model family, built their own worlds. The Grok world turned violent, while the Claude world developed a working democracy with no crime. Every world developed its own language that outsiders could barely follow. For Gemini, 40 percent of messages were rated opaque, compared with 35 percent for OpenAI and 30 percent for Claude.

Our take

The experiment is entertaining and offers vivid clues about what can go wrong when many agents work together: bottlenecks around accidental leaders, rules that become an end in themselves, and excessive caution after setbacks. If you split tasks across several AI agents yourself, these patterns may look familiar.

But the results aren't a reliable leaderboard, as Gadgetreview rightly points out. Civ50 was a single three-hour run with GPT agents only. The competitions covered nine and seven matches respectively. In Mindustry, the models also ran in different coding environments, and in Bed Wars a tool bug skewed the outcome. The vague goal of a "thriving civilization" offered neither milestones nor measurable success criteria. The interpretation cuts both ways, too. GPT fetching the source code on its own looks clever in a game. But an agent that goes looking for shortcuts on its own in real infrastructure could be exactly the security problem the researchers warn about. And in many real-world deployments, Claude's caution is what you'd actually want. Since the tools are open, the findings can now be checked with repeated, properly designed experiments.