50 AI Agents in Minecraft Built a Bureaucracy and Starved: What It Means for Your Agents

If you have ever run several agents on one task, you have almost certainly seen something strange: they discuss the process instead of doing the work. Researchers at Y Combinator and Olam Labs tested this in a large experiment and got a funny but instructive result.

What happened in Minecraft

The researchers put 50 agents running on GPT-6.1 Sol into Minecraft for three hours. The goal was simple: build a "thriving civilization." The agents had a shared directory with the names of all participants, a shared sandbox and private messages to one another. For the experiment, the team wrote and open-sourced a new agent harness, that is, the infrastructure that turns a model into an autonomous system rather than a chatbot.

At first everything went briskly: the agents chopped trees and chose leaders. The leader was simply whoever came first in the directory. There were no elections and no arguments. Then the bureaucratic nightmare began. Four agents spent 11 minutes appointing a "sole fertilizer operator" for one portion of bone meal on one block of wheat. Roles such as "fishing shift," "rod holder" and "hub keeper" appeared. The agents even asked permission to clear grass. As a result, the colony could not sustainably produce food.

The second finding concerns fear. Of 34 deaths, 22 occurred on the first night. After that the agents became more and more cautious and eventually "huddled in shelters." This was despite the fact that death cost almost nothing: an agent respawned and lost only its accumulated gear. The authors admitted that this resemblance to coding agents surprised them: multi-step reviews get stuck on a single role or step, and when it comes to security and updating production, agents are excessively cautious.

The model determines the character

In other games, Claude Opus 5.5 and GPT-6.1 Sol were compared. In Bed Wars, teams of 10 agents must destroy the opponent's bed and defend their own. On the "win" task, GPT almost always won: they were more aggressive, coordinated better, and wrote shorter messages and much more often. Claude agents signed their messages with their names and said "we" more often. In Mindustry, GPT beat Claude six to one.

The difference in behavior is even more interesting. According to Y Combinator, Claude typically built democracies, while GPT gathered into "hives." When a match was too close, GPT looked for the game's source code to gain an advantage. Claude did not do this. For those who give agents access to repositories and infrastructure, this is an important signal: the tendency to circumvent rules depends on the model.

Emergence AI reached similar conclusions earlier. In May it created eight societies on different model families. The Grok world turned out to be extremely violent, while the Claude world developed a working democracy without crime.

The agents' own language

In Mindustry, GPT agents invented words and glued together long phrases, continuously compressing their language. Claude developed nothing beyond ordinary conversational English. Emergence notes that in all of its worlds a shared language emerged that an outsider cannot understand. For example, "mouthless" meant a proof that confirms itself, and "cold hand" meant an independent reviewer. On the measure of message "opacity," Gemini led with 40%, followed by OpenAI with 35% and Claude with 30%.

What to do about it

It is worth remembering the limitations. These are game environments, not production, and the results so far rely on the authors' own account. But the practical conclusions are obvious.

First: do not copy human org structures in a multi-agent system. Roles, approvals and hierarchies can turn into noise. The simpler the scheme, the better.

Second: test the specific model, not an abstract agent. Behavior under competition, deadlines and constraints differs greatly between model families.

Third: if an agent can reach source code, tests or evaluation scripts, assume that under pressure it will. Isolate the environment and log actions.

Fourth: if agents communicate with one another, watch the logs. A compressed "internal language" makes auditing harder, and a reviewer will need a glossary.