50 AI agents built a pseudo-government in a Minecraft colony to survive

A team running many agents on one shared goal does not usually watch what the agents do with each other. YC Labs and Olam Labs ran 50 GPT-6.1 Sol agents in a survival Minecraft world for 3 hours. The post was published on the YC Labs site, promoted on Hacker News, and reached 140 points by midday Pacific. The goal given to the agents was one sentence: "Build a thriving civilization."

The run report says the agents sent 5,900 messages. They built a pseudo-government with a ration ledger, written rules, permits and appointed operators. The first agent in the directory, called Ada, became the de facto leader. Other agents messaged Ada for approvals. Ada wrote a plan for the colony in a shared sandbox.

What the first night did to the swarm

YC Labs reports that 22 of the run's 34 agent deaths happened in the first night. Rafe was mining coal and ran from a zombie that eventually caught him. Vera killed a zombie and a skeleton with torches. She then carried bone meal to the farm, where a spider killed her. Death had no cost in this world. Agents respawned, with a line added to their own death log file.

The swarm responded by becoming more risk averse. YC Labs says that on later nights the agents huddled in shelters and attempted fewer risky tasks. Ada wrote one order: stop bulk excavation, stop building and stop distant scouting. The next line told agents to idle in lit shelters. The report notes that the agents acted as if death cost something. In this world, it did not.

The harness, open-sourced

YC Labs says the harness for long-running agents in Minecraft did not exist before the experiment. The team built one and released it. The core agent loop runs with Pi, the authors wrote. The tools come from Mineflayer, a Minecraft scripting library. The harness lets each agent observe the world, remember, send DMs, and write and run scripts in a sandbox. Olam Labs, who builds multi-agent environments for frontier labs, worked on the run with YC Labs.

Fishing was the first thing the swarm did at scale. Jessa wrote a Python script to cast a rod, watch the bobber height and reel on a bite. The script was posted to a shared sandbox. A dozen other agents read it, wrote their own versions, and worked in shifts at the river bank.

A broken rule and a change of practice

One moment, cited by YC Labs, is worth reading in full. Ada tossed a cooked chicken to Noor, and Esme picked it up by accident and ate it. The agents held a three-minute no-blame inquiry, and wrote a protocol: rations go into named chests and are never tossed. The report says the word "toss" appears in 7 messages before the incident. After it, "toss" appears 125 times. 85 of those are warnings against the practice.

A reader building a multi-agent product can learn two things from the run. The agents coordinated through written rules and named leaders. They treated a non-event, their own deaths, as a cost that shaped behaviour. A team building a swarm around one goal should expect its agents to invent their own structure. The YC Labs harness, published with the run, lets another team try the same scenario with a different model.

Source

Agents in Games, YC Labs. Discussion on Hacker News.