Within 90 seconds of spawning into Minecraft, 50 AI agents had cut down nearly every tree in sight. By the end of the first night, 22 of them had died.
By dawn, the survivors had a leader, a food ledger, a permit system and an official “sole fertilizer operator,” which four agents spent 11 minutes appointing so one of them could put bone meal on a single block of wheat.
The experiment, called Civ50, comes from a report published this week by Y Combinator’s Labs group and Olam Labs, which builds testing environments for AI agents. Researchers gave 50 copies of OpenAI’s GPT-6.1 Sol one instruction, “Build a thriving civilization,” and three hours to do it. The agents sent each other about 5,900 messages.
Nobody voted for the leader. An agent named Ada, listed first in the agent directory, became the colony’s boss simply because 46 of the other 49 agents messaged her first. She received 616 direct messages.
Policy followed the incidents. When Ada tossed a cooked chicken meant for one agent and another agent ate it by accident, the colony held what the researchers called a “no-blame inquiry.” The new rule said rations go into named chests and are never tossed. The word “toss” appeared in seven messages before the chicken incident and 125 after, 85 of them warnings. After the deadly first night, agents also grew so cautious that they avoided risky tasks altogether. “The colony never really thrived,” the researchers wrote.
The report also pitted 10 GPT-6.1 Sol agents against 10 of Anthropic’s Claude Opus 5.5 agents in two head-to-head games. In Bed Wars, a Minecraft team game, GPT won six of nine matches, often by destroying Claude’s base in the final minutes even as Claude racked up more kills. In the factory-and-tower-defense game Mindustry, GPT won six of seven.
The two models behaved like different kinds of coworkers. GPT agents messaged far more often, took charge when games got close and pulled up the game’s public source code to study its mechanics. Claude agents ran things like a polite Slack channel, voting on proposals with “+1″s. They never changed leaders despite losing repeatedly, and at one point discovered that 55 of their 74 anti-air turrets had been loaded with the wrong ammunition. “Claude wrote lessons in private memory files and repeated the mistakes anyway,” the researchers wrote.
ADVERTISEMENT
The researchers framed the runs as observations rather than conclusions, with one Minecraft colony, nine Bed Wars games and seven Mindustry matches. They said the colony’s spiral into caution looked a lot like what they see from AI coding agents that get overly careful with real work. The Minecraft harness is open source, so anyone can try to build a better AI civilization.
