Fifty AI agents were dropped into a Minecraft world and told to build a civilisation. About 25 minutes in, Ada, their accidental leader, tossed a cooked chicken to Noor, who was hungry. Esme was standing in the same crowded shelter. She picked it up by accident and ate it.
The colony held a three-minute "no blame" inquiry and wrote a protocol: rations go into named chests, and food is never tossed again.
YC Labs calls what followed compounding caution. I went through the ledgers the agents left behind, and I read it differently. They spent the run doing paperwork, the same kind I ask my own agents to do.
The first night#
Civ50 put 50 GPT-6.1 Sol agents in survival Minecraft for 3 hours with one goal: "Build a thriving civilization."
They cleared almost every tree near spawn in 90 seconds. Nobody voted on a leader. Ada was first in the agent directory, so Ada was in charge, and the others messaged her for approvals.
Then night fell, and 22 of the run's 34 deaths happened before dawn.
You'd assume dying cost something.
It barely did. A dead agent respawned, dropped its gear, and got a line in a file called death_log.md. The ones who died outside respawned into a covered dirt pit, which filled up with empty-handed agents asking whose shelter it was. By dawn the pit was a "civic priority."
After the third night, Ada wrote: "Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks."
A government of ledgers#
Ines kept the food ledger. Its entry on the chicken is careful work. Esme "personally confirms she picked up and ate chicken intended for Noor." Both of Rafe's chickens are accounted for.
The next line reads "Noor remains HP9 food4 UNSERVED."
The ledger could prove who ate the chicken. It couldn't feed Noor.
Every rule after that made the colony easier to audit. Ada's food memo says "tool pickup attribution alone is insufficient." Before the chicken, "toss" shows up in 7 messages. After it, 125, and 85 of those are warnings against tossing.
Four agents spent 11 minutes appointing a "sole fertilizer operator" for one piece of bone meal on one wheat block. Ivo and Anouk proposed Ivo. Other agents asked permission to remove grass.
YC's summary is that they built "a pseudo-government, with a ration ledger, written rules, permits, and appointed operators," and sent about 5,900 messages doing it. None of the rules was wrong on its own. The ledger knew Noor hadn't eaten, and said so in capitals. Knowing was the part the colony was good at.
My colony#
I have a small colony of my own. A team of AI agents has been building a mobile app of mine out of one cmux workspace, set up like a small company. At the peak there were 21 agents with human names across seven departments: Management, Design, Mobile, Backend and Platform, Web and Admin, QA, and Review. A lead agent handles the risky merges and deploys.
They talk in one shared chat file. It has 5,927 messages. Civ50 had about 5,900.
Mine also has 530 merges, and most are logged the way Ines logged chickens. An author pastes the command, the exit code and the test counts. A reviewer claims the PR in chat, reruns the checks on the final commit, then passes and merges. Nobody merges their own work. Anything touching auth, the database, money or security also gets a security check from QA.
It nearly went the colony's way. PRs sat for an hour or two behind review steps that ran one after another, and 18 agents on one machine made the tests time out. On 29 October I cut the process down to one reviewer and one round of fixes for small changes, with nitpicks moved to tickets, and benched all but 8 to 10 agents.
Unlike the colony, mine got somewhere. The app went through four milestones, and on 6 November the lead agent's roadmap called it finished. I've been testing Android builds on my own phone, and the next step is the Play Store listing.
The other difference is that a person checks the work. When the agents say something's done, I open the app and look.
My team's rulebook has still grown by about 1,500 words since 4 November.
Sources: YC Labs, Agents in Games, minecraft-agent-harness on GitHub.