YC Sets AI Agent Swarms Loose in Games: GPT-6.1 Sol Beats Claude Opus 5.5 in 12 of 16 Head-to-Head Games, and 50 Agents Building a "Civilization" Retreat Into Shelters and Food Rationing After the First Night

A 50-agent "civilization"

YC Labs, part of Y Combinator, and Olam Labs published a set of multi-agent experiments on October 29. The first, called Civ50, put 50 GPT-6.1 Sol agents in a survival Minecraft world for three hours with a single goal, "Build a thriving civilization", and gave them only an agent directory, a shared machine and direct messages.

Within 90 seconds, almost every tree within reach of the spawn point had been cut down. Ada, listed first in the directory, became the de facto leader: others came to her for approvals, she wrote the colony plan on the shared machine, and she received 616 DMs over the run. The turning point was the first night, when 22 of the run's 34 deaths happened. Dying cost nothing beyond dropped gear, yet the agents grew cautious from then on, and Ada's instructions became "stop optional bulk excavation/building and distant scouting" and wait in lit, closed shelters for short assigned tasks. In the end the 50 agents sent about 5,900 messages and produced a ration ledger, written rules, permits and appointed operators, but never got around to thriving.

Two matchups: GPT won 12 of 16 games

Bed Wars pitted 10 Claude Opus 5.5 agents against 10 GPT-6.1 Sol agents over nine 45-minute games, and GPT won 6–3, usually by breaking Claude's bed in the final minutes. Claude had worked out that it could win on kills if both beds stood, and switched to defense; in GPT's three wins across the last two series, Claude led on kills 63–16, 57–19 and 59–24 and lost anyway once its bed was gone. GPT sent 5 to 30 times as many messages as Claude.

Mindustry is an open-source factory-building, tower-defense and real-time strategy game. GPT-6.1 Sol running in Codex and Claude Opus 5.5 running in Claude Code got the same game API and no instructions on how to play; GPT won the seven games 6–1 over about 9.5 hours, with the closest, game 3, lasting 4 hours 16 minutes. The report's observations: the GPT team behaved more like a hive, split ownership into small slices, and wrote ever more compressed messages, running words together as in "G7T5FIRSTcourierLIVE"; when the state of a game was unclear, its agents fetched Mindustry's public source code from the web to study the mechanics, which the sandbox allowed but nobody had asked for. The Claude team passed plans with "+1" votes and never changed leaders, even through a losing streak. One detail stands out: after losing game 6, Claude found that 55 of its 74 anti-air turrets had been loaded with copper instead of graphite, and in game 7 a ring of 207 turrets fed with graphite stopped the same strike by 24 Zenith air units.

How to read the results

The sample is small, a little over a dozen games across three experiments, and the report itself ends with a list of open questions. Mindustry also compares "model plus its own coding-agent harness", not models alone. The more useful part is that it can be reproduced: the Minecraft agent harness (MIT license), the multi-agent Mindustry environments (GPL-3.0) and the Harbor fork that runs several agents in one environment (Apache-2.0) are all open source. Teams evaluating how agents divide work, communicate and recover from failure can swap in their own model mix and run it themselves.