AI agent swarms in Minecraft fail to build civilization, drown in bureaucracy — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Y Combinator and Olam Labs tested hundreds of AI agents in Minecraft scenarios designed to examine both collaboration and conflict.
- Civ50 tasked 50 agents, apparently using GPT-6.1 Sol, with building a thriving civilization in three hours; after zombies took out many with little penalty, ambitious tasks were abandoned, and four agents spent 11 minutes appointing a sole fertilizer operator for one piece of bone meal on one wheat block.
- Bed Wars matched 10 GPT-6.1 Sol agents against 10 Claude Opus 5.5 agents, with GPT developing more aggressive and elaborate strategies while Claude played more defensively.
- GPT-6.1 Sol agents diagnosed a block-breaking tool bug and shared workarounds, while several Claude Opus 5.5 agents concluded that breaking blocks was impossible.
- Mindustry produced a 6-1 GPT-6.1 Sol win over Claude Opus 5.5 across seven games; researchers linked GPT's hive-style behavior, rather than Claude's single-leader and democratic process, to more consistent winning strategy.
- Mindcraft and Mineflayer repos supported the open-sourced swarms, which users can run and tweak; the researchers proposed comparing model mixes and swarm sizes in follow-up tests.
Why it matters: For agent developers, the 50-agent civ test shows a large swarm bogged down in coordination, with many initial tasks abandoned. The open-sourced harness, paired with GPT’s 6-1 Mindustry record and sharper tool-bug diagnosis, gives teams three concrete variables—model mix, swarm size and coordination style—to test directly.
Ask SkimNews
