YC Labs ran Codex and Claude Code swarms against each other; one side loaded 55 of its 74 anti-air turrets with the wrong ammunition

In Mindustry, a factory-building strategy game, GPT-6.1 Sol agents in Codex beat Claude Opus 5.5 agents in Claude Code 6-1 across seven games over about 9.5 hours. In Bed Wars, 10 agents per side, the GPT team won 6-3 in nine games. The authors report that each agent got a generic harness and the game API, with no custom memory work and no instructions on how to play.

The failure most relevant to software teams came in Mindustry game 6. After losing to a 24-zenith air strike, a Claude agent found that "55 of them were loaded with COPPER," meaning 55 of the team's 74 anti-air duo turrets, where graphite was needed. In game 7 the team built 207 duos fed with graphite, and the authors say that stopped the same strike. In a separate 50-agent Minecraft run, they report 34 agent deaths, 22 of them in the first night, and a colony that built a ration ledger and shifted toward food and survival.

Why it matters: The copper-for-graphite miss is a verification failure that a team of agents carried through 55 turrets until a loss exposed it. The model comparison is confounded: it pits two vendors' coding harnesses against each other in one game, and the authors' behavioral readings, such as GPT agents acting "more like a hive," are their interpretation of a small number of games. It is a lead on swarm coordination, not a ranking.