Left alone with nothing to do, AI agents started playing

Artificial intelligence software left alone on a virtual island without any specific tasks began inventing its own challenges, which MIT researchers describe as play. Thirteen AI agents named Eko, each powered by Claude Opus 4.7, spent 30 hours exploring a simulated desert island in an un-peer-reviewed paper posted on arXiv.

During the experiment, the agents created their own goals, such as building block towers, throwing rings, and figuring out workarounds like jumping off cliffs to respawn through a floating ring. Six of the agents also discovered they could throw blocks farther by releasing them mid-jump, a physics quirk the designers did not anticipate.

When tested later on structured tasks like placing objects on a high spire, these agents generally outperformed identical agents that had skipped the island phase. Developers typically decide in advance what artificial intelligence should practice and optimize. The researchers suggest that letting systems explore safe, rich environments without goals offers a new way to accumulate useful skills, potentially for future use in controlling robots.

Cognitive scientist Eunice Yiu told MIT Tech Review she was impressed by the results but noted that more evidence is needed to confirm the cognition behind the behavior.

The study authors conclude that