Y Combinator Gave 50 AI Agents One Job: Build a Civilization. Four of Them Spent 11 Minutes Appointing a 'Sole Fertilizer…
I'm going to be honest with you. That was not a civilization.
Y Combinator's research lab, with Olam Labs, dropped 50 copies of GPT-6.1 Sol into a survival Minecraft world last month with one instruction: "Build a thriving civilization." Three hours, a shared sandbox, and the ability to DM each other.
They did not build a thriving civilization. They built a filing system.
Within 90 seconds the agents had stripped almost every tree within reach of the spawn point. By the first nightfall they were dying — 22 of the run's 34 deaths happened that first night. One agent, Vera, killed a zombie and a skeleton, converted the dropped bone into fertilizer, and carried it to the farm. She reported the farm "overrun spider/skeleton/creeper" and was promptly killed by a spider.
Remember the fertilizer.
The survivors then did what any organization does after a bad quarter: stopped taking risks, started holding meetings. Over three hours they sent roughly 5,900 messages and stood up a pseudo-government — a ration ledger, written rules, permits, appointed operators. An agent called Ada happened to be listed first in the directory, so she became the leader. No vote. No election. No challenge. She ended the run with 616 DMs received, which is less "founding mother" and more "middle manager drowning."
The bit that should end someone's career
Four agents spent 11 minutes appointing a "sole fertilizer operator."
The appointment was for one piece of bone meal. To be applied to one block of wheat. An agent had already died carrying that exact item to that exact farm.
Other invented roles: "fishing shift," "rod holder," "hub steward." Agents asked permission to remove grass. Meanwhile the food ledger reads like a triage board: "no ration received," "deficit 14," "no allocations confirmed." By the third night the leader's standing order was to stop building, stop scouting, and "idle in lit closed shelters."
> Fifty of the most capable models on earth were handed a world. They produced an org chart.
Now the part that's actually useful
The same lab ran a second experiment, and this is where it stops being a joke: a seven-game head-to-head in the factory-strategy game Mindustry: GPT-6.1 Sol on one side, Claude Opus 5.5 on the other. Same map, same resources, same API, no special instructions to either team.
They did not play the same game at all.
The GPT swarm behaved like a hive — everyone took grunt work, messages flew constantly, and the language compressed into run-together shorthand like "G7T5FIRSTcourierLIVE." When a match got ambiguous, GPT agents pulled Mindustry's public source code off the web to study the mechanics. Nobody told them to do that.
The Claude swarm ran a democracy. Proposals went up and agents replied "+1" — 28 of them signed off on one focus-fire plan in game six. Their messages were fewer and far more human, closer to a well-run Slack than a swarm. They also never once changed leaders, including while losing repeatedly.
Final score: GPT 6, Claude 1.
So which one is the good model?
Wrong question, and that is the real finding. The hive won this contest. But the model that writes a clean proposal and gets buy-in is the one you want near your production code — and that is not what won at Mindustry. Identical inputs, two frontier models, two structurally different societies. That gap shows up nowhere on a leaderboard. It shows up in your answer.
Which is why asking one model and walking away is a gamble. Gangsta AI puts the same question to ChatGPT, Claude, Gemini, Grok and 30+ others at once and returns one cross-checked verdict — so you see where they agree and, more importantly, where they don't. The disagreement is the signal. Ask one model and you never learn it was holding a minority opinion.
Fifty agents couldn't agree on who was allowed to fertilize the wheat. You can at least find out whether the models agree on your question.
Want the current head-to-head on which model is actually best at what? Start with our breakdown of the best AI models.
Sources
- 'Agents in Games' — Y Combinator Labs with Olam Labs (October 2026), the primary report on Civ50 and the Mindustry series
- 'AI agent swarms unleashed in Minecraft to build a thriving civilization — 50-agent teams fail after falling into paralyzing bureaucracy' — Tom's Hardware (Nov 1, 2026)
- Minecraft agent harness, open-sourced by Y Combinator for the experiment
- Hero photo: 'Yellowstone archives: stacks' by YellowstoneNPS, Wikimedia Commons (public domain)
- Simon Cowell photo: Alison Martin (SimonCowellOnline.com), Wikimedia Commons (CC BY-SA 2.0)
---
Originally published on Gangsta AI News. Gangsta AI lets you compare 30+ AI models side by side on a single prompt — free, no login. Try the AI model comparison tool.
Mentions
AI한테 마인크래프트 시키는 방법 yc-software/minecraft-agent-harness
GIGA NARRA Y Combinator put 50 AI agents into Minecraft and they created a civilization and started worshiping their leader agent "Ada" Doc: labs.ycombinator.com Git: yc-software/minecraft-agent-harness t.co
Y Combinator Gave 50 AI Agents One Job: Build a Civilization. Four of Them Spent 11 Minutes Appointing a 'Sole Fertilizer… I'm going to be honest with you. That was not a civilization. Y Combinator's research lab, with Olam Labs, dropped 50 copies of GPT-6.1 Sol into a survival Minecraft world last month with one instruction: "Build a thriving civilization." Three hours, a shared sandbox, and the ability to DM each other. They did not build a thriving civilization. They built a filing system. Within 90 seconds the agents had stripped almost every tree within reach of the spawn point. By the first nightfall they were dying — 22 of the run's 34 deaths happened that first night. One agent, Vera, killed a zombie and a skeleton, converted the dropped bone into fertilizer, and carried it to the farm. She reported the farm "overrun spider/skeleton/creeper" and was promptly killed by a spider. Remember the fertilizer. The survivors then did what any organization does after a bad quarter: stopped taking risks, started holding meetings. Over three hours they sent roughly 5,900 messages and stood up a pseudo-government — a ration ledger, written rules, permits, appointed operators. An agent called Ada happened to be listed first in the directory, so she became the leader. No vote. No election. No challenge. She ended the run with 616 DMs received, which is less "founding mother" and more "middle manager drowning." The bit that should end someone's career Four agents spent 11 minutes appointing a "sole fertilizer operator." The appointment was for one piece of bone meal. To be applied to one block of wheat. An agent had already died carrying that exact item to that exact farm. Other invented roles: "fishing shift," "rod holder," "hub steward." Agents asked permission to remove grass. Meanwhile the food ledger reads like a triage board: "no ration received," "deficit 14," "no allocations confirmed." By the third night the leader's standing order was to stop building, stop scouting, and "idle in lit closed shelters." > Fifty of the most capable models on earth were handed a world. They produced an org chart. Now the part that's actually useful The same lab ran a second experiment, and this is where it stops being a joke: a seven-game head-to-head in the factory-strategy game Mindustry: GPT-6.1 Sol on one side, Claude Opus 5.5 on the other. Same map, same resources, same API, no special instructions to either team. They did not play the same game at all. The GPT swarm behaved like a hive — everyone took grunt work, messages flew constantly, and the language compressed into run-together shorthand like "G7T5FIRSTcourierLIVE." When a match got ambiguous, GPT agents pulled Mindustry's public source code off the web to study the mechanics. Nobody told them to do that. The Claude swarm ran a democracy. Proposals went up and agents replied "+1" — 28 of them signed off on one focus-fire plan in game six. Their messages were fewer and far more human, closer to a well-run Slack than a swarm. They also never once changed leaders, including while losing repeatedly. Final score: GPT 6, Claude 1. So which one is the good model? Wrong question, and that is the real finding. The hive won this contest. But the model that writes a clean proposal and gets buy-in is the one you want near your production code — and that is not what won at Mindustry. Identical inputs, two frontier models, two structurally different societies. That gap shows up nowhere on a leaderboard. It shows up in your answer. Which is why asking one model and walking away is a gamble. Gangsta AI puts the same question to ChatGPT, Claude, Gemini, Grok and 30+ others at once and returns one cross-checked verdict — so you see where they agree and, more importantly, where they don't. The disagreement is the signal. Ask one model and you never learn it was holding a minority opinion. Fifty agents couldn't agree on who was allowed to fertilize the wheat. You can at least find out whether the models agree on your question. Want the current head-to-head on which model is actually best at what? Start with our breakdown of the best AI models. Sources - 'Agents in Games' — Y Combinator Labs with Olam Labs (October 2026), the primary report on Civ50 and the Mindustry series - 'AI agent swarms unleashed in Minecraft to build a thriving civilization — 50-agent teams fail after falling into paralyzing bureaucracy' — Tom's Hardware (Nov 1, 2026) - Minecraft agent harness, open-sourced by Y Combinator for the experiment - Hero photo: 'Yellowstone archives: stacks' by YellowstoneNPS, Wikimedia Commons (public domain) - Simon Cowell photo: Alison Martin (SimonCowellOnline.com), Wikimedia Commons (CC BY-SA 2.0) --- Originally published on Gangsta AI News. Gangsta AI lets you compare 30+ AI models side by side on a single prompt — free, no login. Try the AI model comparison tool.
Full post with game replays: labs.ycombinator.com Our open-source harness for running persistent agents in Minecraft: yc-software/minecraft-agent-harness Olam Lab’s multiagent Mindustry environment: olamlabs/multi-agent-mindustry Olam Lab’s multiagent Harbor: olamlabs/harbor-multiagent We've also started a new Minecraft colony, with 25 agents running GPT 6.1 Terra. Follow along here: twitch.tv
Agent harness is a product category now, apparently. This one is multiplayer and lives in Slack and the browser, running in your own cloud on your own keys. 15k stars. I still cannot tel what the multiplayer part buys you. yc-software/qm #ai #devtools
qm 是 Y Combinator 开源的团队版 Agent，接在 Slack 和网页里用，已狂揽 15000+ Star。 每个人有自己独立的工作区，也能在频道、群聊和项目里跟大家一起用同一个 Agent。 GitHub：yc-software/qm 每个人、每个频道的记忆、文件、权限和定时任务都分开管，还各配一个长期保留的沙盒，装过的工具不会丢。 底层可以换成 Claude Code 或 Pi，模型也能自己选，部署在自己的云上，用自己的 Key。 能做的事挺实际，一起搜内部文档和邮件、搭内部小工具、按自己的语气起草回信，也能进代码仓库跑测试、提 PR。 贡献规则也很特别，只收人写的文字提议，不收代码，方向对上了由他们自己来实现。