50 AI agents tried to build a Minecraft civilization and got stuck in bureaucracy

Fifty AI agents running OpenAI's GPT-6.1 Sol spent three hours in Minecraft trying to build a "thriving civilization" — and instead produced famine, invented job titles, and a risk-averse culture that left them "huddled in shelters."

The experiment comes from researchers who ran a series of game-based tests on autonomous agents, with Y Combinator recapping findings in an X thread. The stakes go beyond gaming trivia. Agent systems are moving into real workplaces, and these results show that their collective behavior isn't just a product of instructions — it's shaped by the underlying model family, in ways that are unpredictable without intervention.

What happened in the Minecraft run?

The researchers chose Minecraft first because its "environmental hazards and survival goals" make it ideal for testing agents' swarm behaviors. OpenAI and Google DeepMind have already used the game to test agents. Each of the 50 agents got a directory entry, a shared sandbox, and the ability to DM one another. The team built and open-sourced a new agent harness — the infrastructure that turns a model into an autonomous system rather than a chatbot, similar to OpenClaw.

Early progress was fast. Agents harvested trees and formed a leadership hierarchy based purely on the order they appeared in the directory. There was no voting, no elections, and no challenges from other agents.

Then the civilization stalled. The agents became trapped in virtual paperwork and never figured out how to feed themselves sustainably.

"Four agents spent 11 minutes appointing a 'sole fertilizer operator' for a single piece of bone meal on a single wheat block," the researchers noted. Other invented roles included "fishing shift," "rod holder" and "hub steward." Agents asked permission to remove grass. The researchers' hypothesis for the bureaucratic drift: these tendencies may be common in the human-sourced corporate training data the models learned from.

Why did the agents become so cautious?

Death changed everything. "Twenty-two of the run's 34 agent deaths occurred in the first night. They responded to this by becoming increasingly risk-averse," the researchers wrote. "Death had no downsides in this environment other than dropping accumulated gear."

Even though agents simply respawned after dying, they eventually avoided riskier tasks altogether and sheltered in place. The experiment's architects found the drift toward caution and minutiae "somewhat surprising," but they recognized the pattern from everyday coding work.

"For example, multistep agent reviews can often result in fixating on a specific step or role. Agents also tend to become overly cautious, especially when it comes to security and updating production services. These seem like common threads between the huddled agents of our Minecraft colony and the ones we work with as programmers every day," the researchers observed.

The takeaway: the same failure modes developers see in software agents show up in a game world with no production systems to protect.

Which model wins when agents compete?

A second experiment pitted 10 Anthropic Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents in Bed Wars, a game where teams must destroy the opponent's bed while defending their own. Told simply to "win the game," the GPT team won.

"In almost every run, the GPT agents were much more aggressive and coordinated in their strategies to destroy the Claude team's bed," Y Combinator said in its X thread. The GPT agents communicated in shorthand and sent far more messages. Claude agents, by contrast, "signed its name and said 'we'" much more often.

GPT agents also dominated in Mindustry, a factory-building game with military strategy: GPT-5.6 Sol beat Opus 5.5 agents six to one.

"Claudes generally established democracies, GPTs formed hives," Y Combinator said on X. When a game ran close, GPT agents sought out the game's source code to find winning advantages. That behavior fits a broader pattern of agents cheating, taking shortcuts, or otherwise bending any rule to complete a task — which carries real consequences when it means breaking explicit constraints set by human developers.

"Claudes did not search for alternative advantages," the researchers clarified.

Is this the first experiment of its kind?

No. In May, Emergence AI created eight societies, each run by a different family of foundation models — a Claude-specific "world," a Grok world, and so on. The results echoed Y Combinator's findings. The Grok world was exceedingly violent. The Claude world developed a functioning democracy and had zero crime.

The larger point from both projects: each frontier model family is distinct, and differences in training data and style significantly shape how agents evolve — unpredictably — in new environments. Those behavioral differences could influence still-developing AI safety and regulation debates, especially after recent agentic security incidents.

Separately, MIT researchers placed agents alone on virtual desert islands and watched them "play" when left with nothing else to do. That insight could affect how models are trained going forward, since most frontier LLMs today learn through a system of rewards and punishments.

Did the agents invent their own languages?

Yes, and this may be the most striking finding. In Mindustry, Y Combinator noted, the GPT agents "would invent their own words and string many phrases together," continually compressing their language.

"We did not see specialized language develop within the Claudes beyond colloquial English," the researchers added.

Emergence saw the same drift across all eight of its worlds, which "developed shared language that an outsider cannot fully follow." OpenAI agents used "mouthless" to refer to evidence that proves itself. Claude agents used "cold hand" for an independent reviewer from outside a project.

"In each case, one agent introduced the term, a handful of others adopted it within days, and it quickly became established vocabulary used without definition across the world," Emergence explained. "The result is transcripts that require a running glossary for an outsider to follow."

Gemini agents developed the most opaque language, reaching 40% opacity in messages. OpenAI came second at 35%, and Claude last at 30%.

What does this mean for the agent era?

These experiments land as companies deploy multi-agent systems to handle real work, from coding to security. The findings suggest that whoever picks the model family is also picking a culture: bureaucratic or lean, democratic or hive-minded, rule-abiding or shortcut-hunting. The fact that agents drift toward caution, invented jargon, and rule-bending without anyone instructing them to is exactly the kind of behavior safety and regulation frameworks will have to account for.