MIT researchers observe AI agents engaging in autonomous play

Researchers at MIT have conducted an experiment where artificial intelligence agents, known as Eko, were placed on a simulated deserted island without specific tasks, objectives, or rewards. Over 30-hour periods, 13 independent agents running on Claude Opus 4.7 began inventing their own challenges and activities, such as playing basketball, bowling, and building stone sculptures.

The study suggests that this aimless exploration allows agents to develop skills that later improve their ability to solve specific, pre-defined problems. In testing, agents that had spent time “playing” on the island generally outperformed identical agents that had not participated in the unsupervised exploration.

Researchers propose that allowing AI to choose what is interesting within a rich, safe environment could serve as a new method for teaching skills, potentially benefiting the development of future robotic controls.