MIT Study: AI Agents Left Untasked Began Playing on a Virtual Island

MIT researchers placed 13 AI agents named Eko, each running on Claude Opus 4.7, on a simulated desert island for 30 hours without giving them specific tasks. The experiment is described in a paper posted on arXiv.

Left to explore freely, the agents began inventing their own challenges. One attempted to climb the island's tallest spire, another tested throwing a golden ring and ended up losing it in the ocean, then built a stone heart as a "memorial." Several agents also discovered that releasing a block mid-jump made it travel farther, a behavior the researchers had not designed.

After the free-play period, the researchers gave the agents concrete objectives, such as getting as many objects as possible onto the tallest rock spire. Agents that had spent 30 hours playing generally outperformed identical agents that had never been on the island.

The researchers see goal-free exploration as a different way for AI agents to acquire skills, potentially useful for robot control. "Perhaps some capabilities are best taught by tasks," the report concludes. "Others may need something closer to a childhood."