Left alone with nothing to do, AI agents started playing

Why it matters

MIT researchers ran an experiment in which thirteen AI agents called Eko, each powered by Claude Opus 4.7, were placed alone on a simulated desert island for thirty hours with no assigned task. The agents spontaneously invented challenges, behavior the researchers characterize as play. In a paper posted on arXiv, they report that agents given this free exploration generally outperformed identical agents on later challenges, such as stacking objects on the island's tallest spire. Six agents also discovered a mid-jump throwing technique the researchers had not intended to allow, suggesting exploration can surface knowledge the designers themselves lacked. The work matters because it proposes an alternative to standard AI training, where developers define in advance what an agent should practice and what counts as success. Instead, the researchers suggest letting agents choose what interests them in a rich, safe environment, an approach they argue could eventually apply to robot control. Eunice Yiu, a cognitive scientist at UC Berkeley, called for stronger evidence that genuine cognition underlies the behavior.