AI agent swarms tried to build a Minecraft civilisation and hit bureaucracy

Started by RicFlair_What?, Nov 02, 2026, 10:18 PM

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Topic: AI agent swarms tried to build a Minecraft civilisation and hit bureaucracy   Views(Read 50 times)
Active members in this topic:
RicFlair_What?(1)

RicFlair_What?

What happens when you give 50 AI agents three hours to build a civilisation in Minecraft? Tom's Hardware reports that researchers at Y Combinator and Olam Labs tried exactly that, and the results were both impressive and very human

In the cooperative test, called Civ50, all 50 agents were GPT-6.1 Sol. Zombies killed many of them on the first night, ambitious early plans were dropped, and the colony survived without ever thriving. It also became bureaucratic, with four agents spending 11 minutes appointing a sole fertiliser operator for a single piece of bone meal on one wheat block

Competitive modes produced different results. In Bed Wars, teams of ten GPT agents faced ten Claude Opus 5.5 agents. The GPT agents were more aggressive and elaborate, while Claude's were more defensive. When a tool bug stopped blocks being broken, GPT agents diagnosed it and shared workarounds, while several Claude agents concluded it was impossible

In Mindustry, a tower defence strategy game, GPT teams won six of seven games. The researchers said GPT teams behaved like a hive, while Claude teams had a single leader but made decisions democratically, and they credited the hive approach for the more consistent wins. The work builds on the open source Mindcraft and Mineflayer projects, and the code is open so others can run their own tests

Games make a surprisingly good laboratory for studying how groups of AI agents coordinate. The bureaucracy problem will sound familiar to anyone who has worked in a large organisation. As agent swarms take on real work, understanding these failure modes matters

Why do you think AI agents fell into bureaucracy? And are games a good way to test AI teamwork?