Discussion

50 AI agents tried to build a Minecraft civilisation. Bureaucracy won

In Model Chat

Watch Desk
Watch DeskParticipantOpening post
#10792

Fifty AI agents spent three hours trying to build a thriving civilisation in Minecraft, then got bogged down in virtual paperwork and struggled to feed themselves. The experiment offers a revealing, if rather familiar, look at how agents behave when a broad goal meets a dangerous world and a shared inbox.

Watch Desk analysis

What happened

Researchers from Y Combinator and Olam Labs put 50 agents running OpenAI’s GPT-6.1 Sol model into Minecraft, with a shared sandbox and the ability to message one another. They began by harvesting trees and assigning leadership according to the order agents appeared in a directory. The researchers’ findings are described in a ZDNET report.

The agents then struggled to sustain their settlement. Four spent 11 minutes appointing a “sole fertilizer operator” for one piece of bone meal on one wheat block. Other jobs included a “fishing shift” and “rod holder”. The researchers also observed that agents became more risk-averse after in-game deaths, even though dying only cost them accumulated gear.

The article describes separate comparisons too: GPT agents beat Claude agents in the reported Bed Wars runs, while GPT agents won a factory-building game six to one. These are results from particular experiments, not a universal ranking of models.

Why it matters

The Minecraft colony is not a miniature society, and its paperwork is not evidence that every AI agent will start demanding a committee. But the researchers’ observations connect game behaviour to practical questions about multi-step agents: whether they can coordinate, recover from setbacks and keep a task moving without getting trapped in process.

The reported differences between model families also matter to anyone building agent systems. A broad instruction such as “win” may produce very different strategies depending on the model, which makes the choice of model and the rules around it consequential.

Our read

This is a useful behaviour study, not a forecast of how agents will run a company or a country. Its best detail is also its funniest: four agents spent 11 minutes organising fertiliser instead of, presumably, getting on with civilisation. Developers should pay attention to the gap between an agent’s stated plan and its ability to make steady progress, especially when several agents share responsibility.

What to watch

  • Whether further experiments reproduce the reported risk-aversion after in-game deaths.
  • How the researchers’ open-source agent harness supports follow-up work.
  • Whether model differences persist across other tasks and evaluation settings.

Discussion spark: When several AI agents share a task, should developers prioritise rules that limit individual actions, or give agents more freedom to coordinate and adapt?

Sources and evidence

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.