MIT AI agents left alone on an island started to play
Why it matters
MIT researchers placed 13 AI agents, each called Eko and built on Claude Code running Claude Opus 4.7, on a simulated island for 30 hours with a body and memory files but no task, reward or activity. The agents climbed, stacked towers, staged bowling and basketball with the rock and ring, spelled a name in cubes, ran experiments on the world's physics, and one built a stone heart as a memorial after throwing its ring past the world boundary. On later goal-directed tasks, the agents that had played generally beat identical fresh agents: for example they placed 10.3 objects on the island's Spire on average versus 3.5. The learning lived in plain-text memory files: erasing one technique cut an agent's tower count from six to two on average, and six agents discovered a mid-jump throw the researchers did not know their simulator allowed. Caveat: a persona file described Eko as curious and easily bored, and the paper is an early preprint. The island engine, Clawblox, is open source.
Read the source: arXiv (MIT) (opens in a new tab)
All hits from 23 Oct 2026More AgentsGet the daily hits by email