Share from 0:00
0:00
/

A Tale of Two AI Agent Minecraft Civilizations

It was the best of times, it was the worst of times. Two AI agent Minecraft civilizations. One built by Y Combinator died. The other thrived. Here’s why.

In 2024, researchers put thirty AI agents into Minecraft and gave them one instruction: “survive with fellow players.” The agents formed roles, including farmers, miners, engineers, and organized themselves into a working settlement.

Share

Two years later, Y Combinator put fifty agents in the same game with a bigger goal: build a thriving civilization. It never thrived. On the first night, 22 agents died, and after that the colony grew increasingly cautious. Agents sheltered in place, attempted fewer and fewer tasks, and the simulation became a bureaucracy.

Why did the 2024 experiment succeed while the other failed? It has everything to do with understanding common multi-agent system failure modes and planning ways to avoid them.

In this week’s video, I walk through both experiments and four things that separated the setup that worked from the one that failed.

Additional Resources

I also provide resources outlining why the Y Combinator experiment failed and the learnings you can take if you’re building or working with multi-agent systems.

Resource: Comparing Two AI Agent Minecraft Civilization Experiments
313KB ∙ PDF file
Download
Download

Resource: Lessons for Multi-Agent Systems Building
304KB ∙ PDF file
Download
Download

This video essay is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This installment is part of the Execute layer, which is about using and building agentic systems that work.

Subscribe to Doing AI Efficiently for AI implementation research, analysis, tips and resources.

Discussion about this video

User's avatar