AI agents in Minecraft built a bureaucracy, not a civilisation
Y Combinator had 50 AI agents spend three hours building a civilisation in Minecraft. They wrote rules, kept a ration ledger and appointed a ‘sole fertilizer operator’.
What happened
Give 50 AI agents a Minecraft world and one instruction, ‘build a thriving civilization’, and what do you get? A ration ledger, written rules, permits and a ‘sole fertilizer operator’.
Y Combinator, the startup investor, ran the experiments with Olam Labs, a company that builds multi-agent environments for frontier AI labs. The write-up, ‘Agents in Games’, went online at the end of October, and the code is open source. In the Civ50 experiment, 50 agents running on OpenAI’s GPT-6.1 Sol played for three hours. They had a directory of all agents, a shared folder of files and could message each other. They did so about 5,900 times.
Within 90 seconds almost every tree near the spawn point had been felled. The first night brought 22 of the run’s 34 deaths, after which the colony grew ever more cautious. Ada, the first name in the directory, became leader without an election: 46 of the other 49 sent their first message to her, and she received 616 in all. Ines kept a ration ledger. When Ada tossed a cooked chicken to Noor and Esme picked it up by accident and ate it, a ‘no blame’ inquiry followed, along with a written rule: food goes into named chests and is never tossed again. And four agents spent 11 minutes appointing a ‘sole fertilizer operator’ for one piece of bone meal on one block of wheat.
Then teams of Anthropic’s Claude Opus 5.5 and GPT-6.1 Sol faced off. In nine games of Bed Wars, where you have to destroy the other side’s bed, GPT won 6–3, usually by breaking Claude’s bed in the final minutes. Claude played defensively and aimed for the most kills, which decide the game if both beds survive. In seven games of the factory and strategy game Mindustry, about nine and a half hours in total, GPT won 6–1. GPT worked as a swarm that sent lots of messages and split the work into small slices; the Claudes agreed proposals with a ‘+1’ and kept the same leader for the whole series.
On Hacker News the write-up drew more than 350 points. Readers who work with agents themselves recognised how quickly they drown in lists and procedures.
Stop optional bulk excavation/building and distant scouting. Idle in lit closed shelters until assigned short productive tasks.
Ada, the colony’s leader, after the third night; from Y Combinator’s report
The background
Agents in a game world are nothing new
In 2023 Stanford researchers put 25 agents in a small town modelled on The Sims. One agent got the idea of throwing a Valentine’s Day party; the others spread the invitations over two days and turned up together on time. Each agent kept a memory, reflected on it and made plans. That same year, Voyager taught itself new skills in Minecraft, stored as code it could reuse. It found 3.3 times as many different items as earlier methods.
In 2024 the company Altera had 10 to more than 1,000 agents live together in Minecraft in Project Sid. They took on their own roles, followed shared rules and changed them, and passed on culture and religion.
Why a game
A game gives agents a world with danger, scarcity and a clear result, and you can repeat it endlessly. Minecraft is familiar ground too: OpenAI and Google DeepMind have used it as a test bed before. The agents play through tools: observing the world, walking, chopping, sending messages and writing their own scripts. Jessa, for one, wrote a fishing script and shared the instructions; a dozen agents then fished in shifts.
What working together costs
Working together mostly costs agents messages, and messages cost computing power. Anthropic wrote in 2025 that agents use about four times as many tokens as an ordinary chat, and multi-agent systems about fifteen times as many. Tokens are the chunks of text you pay for with an AI model. Anthropic says multi-agent systems fit poorly where every agent needs the same knowledge or agents depend heavily on each other.
In Bed Wars, GPT sent 5 to 30 times as many messages as Claude. In one game, the GPT players sent 1,147 and the Claude players 82.
Our analysis
Give agents a vague goal and they start holding meetings
Anyone who has watched a group work without a clear brief will recognise this. ‘Build a thriving civilization’ has no finish line, so they switched to what they could keep track of: not dying, and accounting for everything. The researchers themselves compare it to coding agents that become overly cautious as soon as security or production changes come up. And the leader emerged not from skill but from the order of a list, and with 616 messages she instantly became a bottleneck.
GPT versus Claude is not a ranking of models. In Bed Wars a tool bug stopped blocks from being broken; GPT found a workaround, while several Claudes concluded that breaking was impossible. That partly explains their defensive play. These are a handful of games, and the write-up is so far the only source. The most instructive part is memory. The Claude player Cleo noted after a fall not to chase too far when attacking, then later added ‘Died AGAIN’ underneath, pulled off a bridge by another chase.
Our expectation: companies that deploy swarms of agents will run into exactly these costs, with lots of messages and little work. So give an agent one task with a clear end, limits and a person who decides, and don’t count on a notes file to learn from mistakes.
What stands out
- The leader came from the list In Civ50 Ada became leader because she was at the top. In Bed Wars, each team likewise followed whoever came first in its directory.
- Writing it down is not learning Claude players put lessons in their notes and made the same mistake again. GPT left its notes empty and warned teammates by message.
- GPT read the game’s source code When Mindustry got tense, GPT agents fetched the game’s public code to understand its units and rules. The Claudes never did.
- The language kept getting shorter GPT agents ran words together, as in ‘G7T5FIRSTcourierLIVE’. Even an apology came in that style: ‘Sorryforbad tacticalclaim’.
What it means for you
An agent without a clear task will find itself work. Decide in advance what it does, when it is done and who decides.
- Give one task with a clear end For example: enter today’s invoices into the accounts, then stop. A vague goal produces discussion, not work.
- Choose who decides yourself Don’t let leadership depend on whoever happens to be at the top. A colleague has the final say on what counts.
- Check that a correction sticks An agent that says it will remember something can make the same mistake tomorrow. Test it again after every correction.
Where we come in · Solutions
We build assistants with one task and a person who decides
Our custom AI assistants take on routine work, such as labelling emails, pulling invoice details out of PDFs or entering orders into your ERP. Each assistant gets a clearly defined task, and we agree up front what it does by itself and what a colleague checks first.
What work would you give an agent, and who decides?
Tell us what work you have in mind. We’ll help you think through the task, the limits and who decides, and always reply within one working day.