Idle AI agents taught themselves skills by playing on a virtual island

MIT researchers dropped 13 AI agents, each running on Claude Opus 4.7 and named Eko, onto a simulated desert island and left them for 30 hours with no task. When the researchers later set concrete challenges, like moving as many objects as possible onto the island's tallest rock spire, the agents that had spent time playing generally beat identical agents that had never been there. The study is on arXiv, and TechCrunch covered the results.

What the agents did on their own is the interesting part. One tried to climb the tallest spire, failed, then dragged a rock to a lower summit, stacked two blocks, and jumped across. After it accidentally threw a golden ring into the ocean, it built a stone heart as a memorial; when the ring came back, it used the ring as a basketball hoop and a target, and worked out that stepping off a cliff would respawn it near the ring so it could wear it. Six of the 13 agents discovered, without being told, that letting go of a block mid-jump threw it farther.

The researchers say open-ended play produced knowledge of the environment that the designers themselves did not have. A master "soul" file described Eko as naturally curious. Eunice Yiu, a cognitive scientist at UC Berkeley, was impressed but cautioned that the paper does not prove the models were genuinely playing.

Why it matters

If you build agents, this points to a different training path: let them explore an environment freely instead of optimizing only for set goals, and some skills, perhaps including robot control, may come out that you could not have specified in advance.