Y Combinator experiment: 50 GPT agents tasked with a Minecraft civilisation drift into bureaucracy

What happens when dozens of language-model agents share a world and a vague goal? Researchers at Y Combinator's Labs, working with Olam Labs, a company that builds multi-agent environments for AI developers, tried to find out using video games. Their write-up, titled "Agents in Games" and dated October 2026, describes three experiments and releases the tooling as open source.

Civ50: a colony of permits

In the main test, 50 agents running OpenAI's GPT-6.1 Sol were dropped into survival Minecraft and told to build a thriving civilisation within three hours. They could message one another directly and use a shared agent directory. According to the researchers, the agents exchanged about 5,900 messages and recorded 34 deaths, 22 of them during the first night.

The colony survived, but cautiously. Food shortages pushed it into survival mode, and the agents built what the authors call a pseudo-government with a ration ledger, written rules, permits and appointed operators. In one example, four agents spent 11 minutes appointing a single "fertilizer operator" for one piece of bone meal. The authors say the caution reminded them of patterns they see with coding agents, a comparison that remains anecdotal.

GPT versus Claude

Two other experiments pitted teams against each other. In nine 45-minute games of Bed Wars, 10 GPT-6.1 Sol agents faced 10 Claude Opus 5.5 agents; GPT won the series 6–3, often by breaking the Claude team's bed late in the game, even when Claude led on kills. The GPT team messaged far more and played more aggressively. When a tool bug blocked block-breaking, GPT agents found and shared workarounds, while several Claude agents concluded it could not be done.

In Mindustry, an open-source factory and tower-defence game, GPT-6.1 Sol running in Codex beat Claude Opus 5.5 running in Claude Code 6–1 over seven games. The researchers describe the GPT side as hive-like, with dense communication, while the Claude team kept a single leader throughout.

Limits

These are not peer-reviewed benchmarks. Each setup ran once with a small number of games, the Mindustry teams used different agent harnesses, and the behavioural findings come from reading chat logs. Tom's Hardware, which covered the work, presents it as a source of insight into agent behaviour rather than a general ranking of models. The harnesses are on GitHub, and, according to Tom's Hardware, the authors suggest further tests with other model mixes and swarm sizes.