Researchers at MIT conducted an experiment where 13 AI agents, named Eko and powered by Claude Opus 4.7, were placed individually on a virtual desert island for 30 hours without assigned objectives. During this time, the agents began creating their own challenges and activities, a behavior the researchers liken to play.

After the initial exploration period, the agents were given explicit tasks, such as collecting objects and placing them on the island’s tallest rock spire. Those agents that had spent time freely exploring the island generally outperformed identical agents that had not experienced the unstructured environment.

The study, detailed in a paper posted on arXiv, highlights a potential shift in AI training methods. Traditionally, developers define specific goals and success criteria for AI to practice. In contrast, this experiment allowed agents to autonomously decide what was interesting in a rich, safe environment, enabling them to accumulate useful knowledge through self-directed exploration.

One agent, for example, initially attempted to climb the island’s highest spire but failed. It then devised a backup plan involving moving a rock to a neighboring summit, stacking blocks, and successfully reaching the spire. The agent also experimented with throwing a golden ring, repurposing it as a basketball hoop and even finding a workaround to "wear" the ring by exploiting the game’s respawn mechanics.

These behaviors were not mere curiosities; six agents discovered they could throw objects farther by releasing them mid-jump, an unintended mechanic the researchers had not programmed. This demonstrated that the agents developed knowledge about their environment beyond the designers’ expectations.

The AI agents’ master prompt described them as "naturally curious," encouraging exploration without explicit instructions to invent games or challenges. Eunice Yiu, a cognitive scientist at UC Berkeley, noted the findings are impressive but called for further evidence to confirm the cognitive processes behind the behaviors.

The researchers suggest this approach could complement traditional task-based training, proposing that some AI capabilities might be better developed through open-ended exploration akin to a form of "childhood" learning rather than predefined objectives.