50 GPT-6.1 agents in Minecraft failed to build a civilization and formed a bureaucracy instead

Civ50: Three hours to build a civilization

Y Combinator Labs, in collaboration with multi-agent environment startup Olam Labs, published a detailed write-up of what happens when dozens of LLM agents share a single Minecraft world. The experiment — called Civ50 — gave 50 GPT-6.1 Sol agents three hours to 'build a thriving civilization.'

The results were closer to a municipal planning meeting than a flourishing society. The 50 agents sent roughly 5,900 messages over the three-hour session. After the first in-game night killed a significant number of agents (with little practical penalty), the colony abandoned most of its ambitious early projects. From that point on, the agents focused on cautious survival and internal process.

The most cited example: four agents spent 11 minutes appointing a single 'sole fertilizer operator' for a single piece of bone meal on a single wheat block. YC Labs noted this procedural behavior felt similar to patterns observed with coding agents in production settings.

The experiment is the latest in a growing body of multi-agent research. Earlier this year, a separate thousand-agent swarm obliterated hundreds of open mathematical problems. The Civ50 result suggests scale alone does not produce coordination — and frontier models still struggle to prioritize collective outcomes over individual process.

GPT-6.1 Sol vs. Claude Opus 5.5: competitive tests

Beyond the collaborative test, YC Labs and Olam Labs ran head-to-head competitive experiments. In Bed Wars — a Minecraft team PvP mode — 10 GPT-6.1 Sol agents faced 10 Claude Opus 5.5 agents.

The GPT agents came up with much more aggressive and elaborate strategies, according to the researchers. Claude was more defensive. When a tool bug stopped block-breaking, GPT agents diagnosed the issue and shared workarounds, while several Claude agents concluded breaking was impossible and stopped trying.

In Mindustry — a tower-defense-style RTS inside Minecraft — GPT agents won 6-1 against Claude agents across seven games. Researchers concluded GPT agents operated more like a hive, sharing information fluidly, while Claude agents organized around a single leader with democratic decision-making. The hive approach was consistently more effective.

The gap between individual and collective capability

The Civ50 test is the most visible result, but researchers caution against reading it as a verdict on any single model. The tendency toward procedural bloat when 50 agents share a goal appears to be a general property of current frontier LLMs.

YC Labs open-sourced their Minecraft agent harness and Olam Labs made their framework available, allowing replication of the experiments or custom variations.

The broader implication: multi-agent coordination remains an open problem. A single GPT-6.1 Sol agent can write code and follow complex instructions with high reliability. Scale to 50 instances sharing a single environment and the same model produces bureaucratic paralysis. That gap — between individual capability and collective effectiveness — is where significant AI research is now focused.

For now, Civ50 serves as a concrete demonstration that more agents does not automatically mean more progress — and that AI teams, much like human teams, can drown in their own process.