# Fifty GPT-6.1 Sol agents tried to build a Minecraft civilization and got stuck in procedure

> Researchers from Y Combinator and Olam Labs gave 50 GPT-6.1 Sol agents three hours to build a thriving Minecraft society. Zombies and paperwork won, and head-to-head games against Claude Opus 5.5 followed.
>
> OpenCheese AI · 2026.11.04 · Web page: https://opencheese.dev/blog/fifty-gpt-6-1-sol-agents-tried-to-build-a-minecraft-civilization-and-got-stuck-i

Give fifty AI agents three hours and a mandate to build a prosperous civilization, and what you get, at least in one recent experiment, is mostly process. Researchers from Y Combinator and Olam Labs ran that test in Minecraft with GPT-6.1 Sol. The colony never became prosperous, and its most memorable moment involved a single wheat block.

## What the researchers actually ran

The team launched hundreds of AI agents across Minecraft and other game scenarios. The aim was to watch how autonomous models behave around each other: whether they cooperate, how they split up work, and how they compete when the incentives turn hostile.

They did not build the control layer from scratch. The experiments run on two existing tools, Mindcraft and Mineflayer, which let software agents drive in-game characters.

## Three hours, one night, many meetings

In the first test, 50 agents on GPT-6.1 Sol had 3 hours to build a thriving civilization. They did not manage it.

The first night did most of the damage. Zombies wiped out a significant share of the characters, and many of the original tasks were simply abandoned. The survivors switched to careful survival and, little by little, to negotiating who was responsible for what and by which procedure.

The sharpest example: 4 agents spent 11 minutes deciding who should be the single person in charge of fertilizing one block of wheat with one portion of bone meal. Eleven minutes, four agents, one portion.

The authors drew a comparison to something developers already know. Groups of AI agents working on programming tasks can behave the same way, spending a large part of their resources organizing the work instead of doing it.

## Bed Wars, where the models showed their temperaments

The second test turned cooperation into competition. In Bed Wars mode, a team of 10 GPT-6.1 Sol agents played against a team of 10 Claude Opus 5.5 agents.

The GPT side picked more aggressive and more complicated strategies. The Claude side played more cautiously.

A tool bug exposed another difference. The bug stopped agents from breaking blocks. The GPT agents tracked down the cause and passed workarounds to each other. Several Claude agents concluded that breaking blocks was impossible and acted accordingly.

The source does not say who won Bed Wars, so neither will this article.

## Mindustry and the question of who is in charge

The third scenario was Mindustry, a strategy game with Tower Defense elements. Here the scoreline is on record: the GPT team beat the Claude team in 6 of 7 matches.

The authors' explanation is about structure. GPT agents behaved more like a single collective. Claude's team relied on one leader and a democratic decision-making process. In this setting, the more distributed way of working came out ahead.

There is a wry detail in all this. The model family that lost eleven minutes to bone meal logistics is the same one that took six of seven Mindustry matches. Coordination, it seems, is not free and not useless. The cost depends on what the game is asking for.

| Test | Setup | Reported outcome |
|---|---|---|
| Civilization building | 50 GPT-6.1 Sol agents, 3 hours | Goal not reached, zombies killed many agents, survivors got lost in procedure |
| Bed Wars | 10 GPT-6.1 Sol vs 10 Claude Opus 5.5 | GPT more aggressive, Claude more cautious, GPT agents fixed a tool problem while some Claude agents gave up on block breaking |
| Mindustry | GPT team vs Claude team | GPT won 6 of 7 matches |

## Open code, and a caveat the authors insist on

The researchers published the source code of their experimental systems. Other developers can rerun the tests and check how group size and mixtures of different models change the results.

The authors see games as a way to study how AI agents behave when they have to work together or against each other. They also say plainly that Minecraft results do not prove the same patterns will show up in real working systems. That is a sensible limit, and the experiments are best read as a set of observations rather than a verdict on any model.

Still, anyone who has sat through a meeting about the agenda for another meeting will recognize the bone meal episode at once.

## Related posts

- 2026.10.20 · [Left alone on a virtual island for 30 hours, Claude agents invented games and got better at tasks](https://opencheese.dev/blog/claude-agents-virtual-island-30-hours-play-learning.md): MIT researchers left copies of Claude Opus 4.7 alone on a simulated island with no task. The agents made up games, and the experience later improved their performance. Whether it was really play is disputed.
- 2026.11.04 · [Anthropic tells Claude Code to stop recreating specific existing video games](https://opencheese.dev/blog/anthropic-tells-claude-code-to-stop-recreating-specific-existing-video-games.md): Anthropic added two sentences to Claude Code's prompt on Opus 5.5 so it steers away from cloning named games. Original, genre and own-asset games are still meant to work.
- 2026.10.18 · [AI agents can find the right number in a paper and still miss why the authors used it](https://opencheese.dev/blog/ai-agents-astronomy-reproduction-implicit-knowledge.md): In an astronomy reproduction test, 12 analysis variants gave results from 2.16 to 3.53 kpc against a published value near 2.70. The needed correction was already in the paper.