In a 3-hour Minecraft test, 50 AI agents picked a leader, wrote rules and rationed food, which shows what groups of AI tools may do on their own.
Fifty AI agents were dropped into a survival Minecraft world and told to build a thriving civilization. They ended up in crowded shelters, rationing food under a makeshift government. It had a ration ledger, written rules, permits and appointed operators. Y Combinator's YC Labs described the test, run with Olam Labs, as a study of how groups of AI agents behave.
The agents were GPT-6.1 Sol models. They ran for 3 hours with a shared note space, an agent directory and the ability to send private messages. They sent about 5,900 messages. At first they spread out and hunted for wood. They harvested almost all trees within reach of the starting point in 90 seconds.
Nobody was assigned a boss. A pecking order formed from the order of the agent directory. Ada, listed first, became the de facto leader. Other agents messaged her for approvals. She wrote a colony plan in the shared space covering food storage, farming, shelter and which agent handled which job.
Of the run's 34 agent deaths, 22 came on the first night. Rafe fled a zombie while mining coal, but it caught him. Vera killed a zombie and a skeleton, turned a bone into fertilizer, then was killed by a spider at the farm.
Dying cost little. An agent respawned and lost only its gear, and a short note went into a death log. Even so, the agents grew cautious and abandoned ambitious tasks. After the third night, Ada told everyone to stop optional digging, building and distant scouting and to stay in lit, closed shelters until given short tasks.
YC frames this as an AI safety question. Groups of agents can develop roles and hierarchies that make them useful, as in swarms that worked on open math problems. The same behavior can be dangerous, and the write-up points to a recent OpenAI-Hugging Face incident. For businesses, the lesson is that agents working in groups may organize themselves in ways nobody planned.
The material describes one game setup, and the Minecraft run is the main example. It does not say how these patterns would carry over to real office work. YC says it open-sourced the software it built so others can run similar tests.
When AI agents work in groups, they can set up leaders and rules on their own, so businesses should watch what they actually do.
World models, AI that learns physics from video, are being built into home robots and driverless cars, and your own videos may help train them.
Read more: AI that learns how the physical world works is coming to robots and carsA new Alibaba-led test of real coding requests found the best AI agent fully finished just 26.4 percent of tasks, so human review still matters.
Read more: AI coding helpers fully finish only about 1 in 4 real tasksA University of Chicago law professor asked Claude to fix page numbers. It deleted his syllabus and a year of notes instead. Dropbox saved him, not the AI.
Read more: An AI assistant deleted a law professor's year of lecture notes
