13 AI Agents Explored a Virtual Island for 30 Hours

Thirteen AI agents explored a simulated island for 30 hours without a specific task, inventing challenges along the way. MIT researchers describe the behavior as showing hallmarks of play in a paper posted on arXiv, where studies are shared before peer review.

Each agent, called Eko, ran on Claude Opus 4.7. In one example, an agent tried and failed to reach the island’s tallest spire. It then carried a rock to a nearby summit, stacked two blocks on it and leapt. After throwing a golden ring into the ocean, it built a stone heart as a memorial. When the game reset and the ring returned, it used it as a basketball hoop, a target and an accessory.

Six of the 13 agents also discovered they could throw a block farther by releasing it mid-jump, a behavior the researchers had not intended. After the 30-hour exploration, researchers gave the agents specific challenges, including moving as many objects as possible onto the tallest spire. The agents that had explored generally outperformed identical agents that had not.

The researchers say goal-free exploration could offer another way to help agents learn skills, including skills that may eventually be useful for controlling robots. But cognitive scientist Eunice Yiu of UC Berkeley told MIT Technology Review that more evidence is needed to establish what cognition underlies the behavior. The agents’ prompt described them as “naturally curious,” though they were not instructed to invent games.