MIT researchers placed 13 AI agents named Eko, each running on Claude Opus 4.7, alone on a simulated desert island and let them explore without explicit objectives for 30 hours. The team describes the experiment in a paper posted on arXiv. After the unguided exploration period, the researchers assigned concrete tasks — for instance, getting as many objects as possible onto the island’s tallest rock spire — and found that the Eko agents that had spent those 30 hours exploring generally outperformed identical agents that had never been on the island.
What the agents did
The Eko agents were not instructed to invent games, but their master prompt or “soul” file characterized them as “naturally curious,” which biased their behavior. In its first hour, one Eko wrote, "Let me observe my surroundings," and ambitiously attempted to conquer the island’s highest spire. After failing, it returned with a backup plan: it hauled a rock to a neighboring summit, stacked two blocks on top of it, and leapt up, later announcing, "I’M ON THE SPIRE."
Another episode involved an arcing experiment with a golden ring. The agent tested how far it could throw the ring, only to send it into the ocean: "Ouch. Lost the Ring." It then decided to "pivot to something achievable and meaningful" and built a heart from stones as "a memorial to the lost Ring." When the game reset and the ring reappeared, the agent reused it as a basketball hoop, a target, and eventually an accessory. When the environment would not allow placing the ring around its body, the agent discovered a workaround: falling off the cliff would respawn it in the same area as the ring, and it deliberately stepped off and reappeared through the ring, later stating, "I’m wearing the Ring."
These improvisations had practical consequences as well. Six of the 13 agents discovered they could throw a block farther by releasing it mid-jump — a mechanic the researchers had not intended to be possible. The paper characterizes the exploration as producing "knowledge of the world that its designers lacked."
Why this matters
Today, developers typically specify in advance what AI should practice and how progress is measured. The researchers propose that giving an agent a sufficiently rich and safe environment and letting it choose what is interesting amounts to a different learning regime. Such unguided, play-like exploration could accumulate useful knowledge that later helps performance on directed tasks, and could be especially relevant for teaching skills needed to control robots.
The team notes that some capabilities may indeed be best taught by explicit tasks, while others might benefit from something closer to a childhood—unstructured time to explore and experiment.
Expert reaction and limits
Eunice Yiu, a cognitive scientist at the University of California, Berkeley, told MIT Tech Review she was impressed but wants stronger evidence about the cognition behind the behavior: "It could be the case that the models are genuinely playing, but we don’t know based on the paper," she said. The study is a preprint on arXiv and has not been peer-reviewed at the time of posting, which the authors and observers treat as a reason for further validation.
Conclusion
The experiment shows that minimal constraints plus a characterization of curiosity can lead AI agents to invent challenges and acquire practical tricks that their designers did not foresee. The researchers suggest this kind of open-ended exploration can complement task-based training, but they also call for more work to understand when and how such play-like behavior yields robust, generalizable capabilities.
