Researchers from MIT left 13 AI agents named Eko alone in a simulated island environment for 30 hours to study what they would do without assigned tasks. Each agent was based on Claude Opus 4.7. The team reported that the agents began creating their own challenges, which the researchers interpreted as play. The study is available on arXiv.
One agent set out to climb the island’s tallest tower but failed. About an hour later, it tried a different approach, stacking blocks on a nearby peak and jumping onto them. In another episode, an agent lost a gold ring in the sea while trying to throw it. It then built a heart out of stones as a memorial before using the ring, after it reappeared when the game was reloaded, as a basketball hoop, a target and an accessory.
The researchers had not directly instructed the agents to invent games. However, the master prompt described Eko as naturally curious, and the environment was designed to encourage human-like behavior. The team also found that play led some agents to develop new skills: six of the 13 learned to throw a block farther by releasing it in mid-jump. The researchers said this revealed knowledge about the world that even the developers had not anticipated.
After the 30-hour period, the researchers gave the agents specific tasks, including moving as many objects as possible to the island’s highest peak. Agents that had spent time exploring and playing generally outperformed otherwise identical agents with no prior experience in the environment.
The researchers suggest that unguided exploration could offer another way for AI agents to learn, particularly for future robot applications. Rather than deciding in advance what an agent should practise or optimise, they propose giving it a varied, safe environment in which it can choose what to explore and learn from the resulting experiences. Some abilities, they argue, may be better taught through tasks, while others could benefit from something closer to a childhood.
Sources: t3n.