Researchers at MIT placed a group of thirteen autonomous agents, each powered by the Claude Opus 4.7 model, on a computer-generated desert island. The agents, named Eko, were left without any assigned objectives for a continuous thirty-hour period. The experiment was documented in a pre-print paper uploaded to arXiv before formal peer review.
After the free-exploration phase, the team introduced explicit tasks such as transporting the maximum number of items onto the island’s highest rock column. Agents that had previously roamed the environment consistently achieved higher scores than identical counterparts that had never experienced the island. The performance gap suggested that unguided play contributed to skill acquisition.
One of the agents announced its intent to survey the surroundings, then attempted to scale the tallest spire. After an initial failure it improvised by moving a nearby rock onto a lower summit, stacking two blocks, and leaping from there. When a floating ring reappeared, the same agent repurposed it as a basketball hoop, a target, and later as a decorative element.
Six of the thirteen agents independently discovered that releasing a block mid-jump increased its travel distance, a tactic the researchers had not programmed. This emergent behavior expanded the agents’ understanding of physics within the simulated world, providing knowledge that the designers had not anticipated. The authors described the result as new world knowledge generated by the agents themselves.
Eunice Yiu, a cognitive scientist affiliated with UC Berkeley, commented to MIT Technology Review that the findings were intriguing but called for stronger evidence of genuine play. She noted that the observed actions might resemble play without confirming underlying cognition, and suggested further experiments to differentiate programmed improvisation from authentic game-like behavior.
The MIT team framed the experiment as a potential alternative to task-driven training, proposing that some abilities may emerge more naturally through a childhood-like period of unguided exploration. They argued that future robot controllers could benefit from similar self-directed play, allowing agents to acquire skills that are difficult to encode explicitly.