Claudes built democracies, GPTs formed hives: what YC learned running AI agent swarms in Minecraft and Mindustry
What happened? Y Combinator's research group YC Labs, working with Olam LabsDealroom has a profile for this one. Try Dealroom →, a startup that builds multi-agent environments for frontier AI labs, has published the results of letting swarms of AI agents share a game world for hours. It ran three experiments: a 50-agent Minecraft colony, a Minecraft Bed Wars series, and a series of the factory and tower-defense game Mindustry, pitting OpenAI's GPT-6.1 Sol against Anthropic's Claude Opus 5.5.
The scoreline: GPT won both head-to-heads: 6-3 over nine 45-minute Bed Wars games (10 agents a side) and 6-1 over seven Mindustry games across about 9.5 hours, where GPT ran inside Codex and Claude inside Claude Code with the same game API.
Democracies vs hives: The Claude teams kept a single leader and decided plans by posting proposals that teammates approved with a "+1", and never switched leaders despite losing many games in a row. GPT agents split ownership into small slices, shared resources freely and messaged 5 to 30 times more often, their chat compressing into strings like "G7T5FIRSTcourierLIVE". When games got tight, GPT agents fetched Mindustry's public source code from the web to understand the mechanics; the Claudes never did. In Bed Wars, a tool bug stopped block-breaking: a GPT agent worked out a workaround and shared it, while several Claudes concluded breaking was impossible and kept guarding.
Accidental bureaucracy: In the 50-agent colony, told only to "build a thriving civilization" over three hours, most agents sent their first message to Ada, simply the first name on the roster. She received 616 DMs and became the leader with no vote. After 22 of 34 deaths came on the first night, the colony grew steadily more cautious, exchanged about 5,900 messages, spent 11 minutes appointing a "sole fertilizer operator" for one piece of bone meal, and ran a three-minute "no blame" inquiry after one agent ate a chicken meant for another.
Why it matters: As companies start deploying fleets of agents, the experiments show emergent dynamics nobody designed: roster order deciding who coordinates, leaders that never get revisited, and shared state nobody checks (one Claude found 55 of its team's 74 anti-air turrets loaded with the wrong ammunition). The authors frame it as alignment research. The runs are small, single-configuration and self-reported, with model and harness tangled together in Mindustry, so they don't rank models. The environments are open source for others to rerun.
Sources: Y Combinator on X (YC Labs write-up), RuntimeWire, Hacker News discussion
Image credit: YC Labs / Olam Labs
