AI agents allowed to explore a simulated island without a fixed objective invented challenges and later performed better on assigned tasks than otherwise identical agents without that experience, according to an MIT study posted to arXiv.
Researchers placed 13 agents called Eko, each based on Claude Opus 4.7, alone in the environment for 30 hours. They climbed structures, threw objects and discovered unintended mechanics, including a way to launch a block farther by releasing it during a jump. After the free-exploration period, the agents were asked to complete goals such as moving as many objects as possible onto a tall rock spire.
The previously exposed agents generally outperformed the control agents. Researchers argue that open-ended play could help systems acquire useful knowledge without developers defining every exercise and reward in advance. That may matter for embodied agents and robots operating in environments whose useful skills are hard to enumerate beforehand.
The interpretation remains uncertain. The agents’ system prompt described them as naturally curious, and an outside cognitive scientist said the observed behavior does not establish that the models were genuinely playing. The paper has also not completed peer review. The result supports free exploration as a training method, not a claim that current agents experience curiosity in the human sense.