50 AI agents in Minecraft ended up building a bureaucracy: without a success criterion, autonomy produces paperwork

By El Ecosistema Startup (based on the Y Combinator lab's report) · November 1, 2026.

Y Combinator's research lab and the startup Olam Labs let 50 AI agents loose in a Minecraft survival world with a single instruction: “Build a thriving civilization.” According to the report, three hours later the most reliable thing they had built was a food-ration queue.

All the agents ran on OpenAI's GPT-6.1 Sol. They could send each other private messages, consult a shared directory and write files on a common machine. No one voted for a leader: 46 of the other 49 agents first wrote to Ada, the first on the list, who ended up receiving 616 direct messages.

On the first night, monsters caused 22 of the game's 34 deaths. Dying cost almost nothing, since an agent respawned instantly, but the colony never regained its nerve. After the third night, Ada asked that no one build optional things or explore distant areas, and that everyone wait in sealed shelters. The researchers described it as a surprising turn toward caution.

Instead, administration grew. One agent, Ines, kept a file of who was hungry and which chest held each fish. Four agents spent 11 minutes appointing a “sole fertilizer officer” to use a single piece of bone meal. About 5,900 messages were sent in three hours. The chicken episode sums it up: one agent accidentally ate a ration thrown to another, there was a “blameless” investigation and a new protocol. The word “throw” went from seven mentions before the incident to 125 after, 85 of them warnings.

**The model showdown.** The lab also pitted GPT against Claude. In nine games of Bed Wars (a Minecraft minigame where each team defends its bed), ten GPT-6.1 Sol agents beat ten Claude Opus 5.5 agents 6 to 3. GPT attacked and sent between five and thirty times more messages. Claude defended and went for kills, and was left exposed to a late-game infiltration. When a tool bug prevented breaking blocks, a GPT agent found a workaround and shared it. Several Claude agents concluded it was impossible.

In Mindustry, a strategy and building game, GPT inside Codex beat Claude inside Claude Code 6 to 1 over about nine and a half hours. The GPT agents downloaded the game's public source code to study its mechanics without anyone asking them to. The Claude agents never tried, though they did learn: after the sixth defeat, one noticed that 55 of 74 anti-air turrets were loaded with copper instead of graphite. Once fixed, the defense stopped the same attack in the seventh game.

On pricing, according to TechRepublic, GPT-6 Sol and Claude Opus 5.5 were released on September 22, 2026, with list prices of $2 per million input tokens for OpenAI versus $4 for Anthropic, and $10 versus $20 for output. According to Unite.ai, the Artificial Analysis index at maximum effort gives Opus 5.5 58 points and GPT-6 Sol 48, with an average cost per task of $5.98 versus $1.06.

**Our take.** The thesis is that bureaucracy is not a glitch in the game: it is what an optimizer does when the perceived cost of error is high and the goal has no metric. “Prosperity” cannot be verified, so the agent optimizes what it can check: not making mistakes. Queues, permissions and logs are the way to be beyond reproach. It shares the same root as *specification gaming*, but in its cautious version: not cheating, but paralysis. And it closely resembles the coding agent that asks for approval at every step or refuses to touch anything that “seems risky.”

The finding fits a pattern we see repeatedly: agent reliability depends not only on the model but on how the task is specified and what oversight layers surround it. Here the oversight was the colony itself, which imposed it on itself. For those deploying agents, the practical takeaway is twofold: define verifiable success criteria before delegating, and measure cost per approved result (first-pass acceptance rate, correction time), not token price. The cost-per-task figures illustrate this: the model with the lowest list price may be neither the most expensive nor the cheapest once retries are counted.

The experiment should be taken with caution, and the article itself acknowledges this. It is a report from an investor's lab, not a peer-reviewed study. The colony was tested only once and only with GPT agents, so we don't know whether a Claude colony would have done anything different. The showdowns compare complete products (each model inside its company's agent software), not isolated models, and a tool bug affected at least two Bed Wars games. That is why no winner should be drawn from a 6-3 or a 6-1.

What does endure is the image. In the short term, delegating to agents without defining what a good result looks like will produce more process than progress, and that also costs time and money. But the fact that the environment and experiments are open source and run on a laptop is a good sign: the way to learn to govern these systems is to test them cheaply and often. That discipline of specifying, measuring and auditing is what will bring us, in the long term, closer to agents that free up work rather than generate paperwork.