MIT stranded an AI agent on a virtual desert island — it started to play

MIT CSAIL placed a Claude-powered agent named Eko on an empty virtual island with no assigned task. Across 13 iterations of 30 hours each, the agents explored, built structures, invented games and set their own goals. The researchers argue that free play could complement task-based training for AI agents.

• 13 versions of agent Eko each spent 30 hours on the island with no tasks
• Agents built shelters, invented games and named a collapsed monument "Hubris"
• Researchers compared the behavior to play in animals and humans
• Conclusion: some AI skills may be better learned through free exploration than tasks