50 AI agents in Minecraft built a government, a ration ledger and a no-blame inquiry about a chicken
YC Labs, working with Olam Labs, put 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with the goal 'Build a thriving civilization'. They sent about 5,900 messages, formed a pseudo-government and ended up cautious and bureaucratic. In a separate Bed Wars test, GPT agents beat Claude agents 6 to 3.
- Agent or tool
- GPT-6.1 Sol agents
- Topic
- Research & Data
- Reported on
- YC Labs
- Added
- 2026-10-30
My write-up
Fifty agents. One Minecraft world. Three hours. The goal: 'Build a thriving civilization.' YC Labs, working with Olam Labs, published what happened, and I'd call it a thriving bureaucracy.
From the write-up: the agents spent the run in crowded shelters, rationing scarce food. They built a pseudo-government with a ration ledger, written rules, permits and appointed operators, and sent about 5,900 messages. There was no election for a leader. 46 of the other 49 agents sent their first message to Ada, the first name on the roster, and she ended up with 616 direct messages. Of the run's 34 agent deaths, 22 were on the first night, and after that the agents got much more risk-averse.
The bureaucracy is the best bit. Four agents spent 11 minutes appointing a 'sole fertilizer operator' for a single piece of bone meal on a single wheat block. Agents asked permission to remove grass. And when Ada tossed a cooked chicken to Noor and Esme picked it up by mistake and ate it, a three-minute 'no blame' inquiry produced a written protocol: rations go in named chests and are never tossed. The word 'toss' appears in 7 messages before the incident and 125 after, 85 of them warnings.
In a separate Bed Wars test, 10 Claude Opus 5.5 agents played 10 GPT 6.1 Sol agents across nine 45-minute games. GPT won 6 to 3, usually by breaking Claude's bed late, while Claude played defensively. Those are the lab's results for one setup, so don't read them as a general ranking.
The authors say the patterns look like what happens with coding agents: fixating on one step or role, and getting overcautious around security and production. I recognise that, and not only in agents.
It got attention fast. When I checked, the Hacker News thread had 248 points and 82 comments. Credit: YC Labs and Olam Labs. The original has replays and the full data.
Why it matters
Groups of agents reproduce familiar workplace problems: hubs, red tape and overcaution. Seeing it in a game makes the risks easy to spot before you build a real multi-agent workflow. It is one of 72 sightings filed under Research & Data on Agent Sightings (number 1 by interest), and one of 1 involving GPT-6.1 Sol agents.
How you could try it
Here is a reusable prompt, written by us in our own words, that captures the task. Paste it into GPT-6.1 Sol agents or any similar agent and adjust the details to your situation.
I'm designing a workflow with several AI agents. Review it for the failure modes seen in agent colonies. Check for a single hub that everyone messages, so it becomes a bottleneck; for agents spending effort on process, such as approvals, roles and permission requests, rather than the task; for handoffs where one agent passes something and another picks it up by mistake; and for agents becoming overcautious after failures. For each risk, propose a concrete fix, such as a named handoff location, a written protocol, a message budget or a deadline. Keep the plan as simple as it can be.
The prompt breaks down into 5 steps you can check as the agent works:
- I'm designing a workflow with several AI agents.
- Review it for the failure modes seen in agent colonies.
- Check for a single hub that everyone messages, so it becomes a bottleneck; for agents spending effort on process, such as approvals, roles and permission requests, rather than the task; for handoffs where one agent passes something and another picks it up by mistake; and for agents becoming overcautious after failures.
- For each risk, propose a concrete fix, such as a named handoff location, a written protocol, a message budget or a deadline.
- Keep the plan as simple as it can be.
Tip: When agents hand things to each other, define the handoff, such as a named place or a written protocol, rather than letting them improvise.
Source
Read the original first-hand account on YC Labs: https://labs.ycombinator.com/projects/games/index.html. Our summary is a paraphrase; details are in the source.
Featured in these guides
Built an AI agent? Feature your agent on Agent Sightings, free submission or a featured spot.