What might an AI agent learn when no one gives it a task? In an experiment described by MIT researchers, agents left on a simulated desert island spent their time exploring, setting challenges and finding unexpected ways to use objects. Their experience also appeared to help when they later faced specific tasks.
From open exploration to a measurable challenge
The researchers put 13 AI agents called Eko on the island for 30 hours. Each ran on Claude Opus 4.7 and was left without a specific task. The study was posted on arXiv, where researchers share work before peer review.
After the exploration period, the researchers gave the agents challenges, including getting as many objects as possible onto the island’s tallest rock spire. The Eko agents that had spent time on the island generally outperformed identical agents that had never been there.
That comparison connects open-ended exploration to a concrete result: performance on a later challenge. The researchers suggest that exploration without a set goal could offer another way for AI agents to develop skills, including skills that might one day be used to control robots.
Agents made their own goals
Eko’s first-hour plan was to observe its surroundings and conquer the tallest spire. After failing to reach it, the agent tried a different route: it carried a rock to a neighboring summit, placed two blocks on top and jumped. It announced, “I’M ON THE SPIRE.”
Another attempt produced a different kind of lesson. Eko tried to throw a golden ring, but sent it into the ocean. After noting that it had lost the ring, the agent shifted to building a heart from stones as a memorial. When the game reset and the ring reappeared, it became a basketball hoop, a target and eventually an accessory.
The game did not allow Eko to put the ring around its body. The agent found a workaround: it stepped off a cliff, respawned near the ring and dropped through its middle. “I’m wearing the Ring,” it announced. These episodes show how the agent’s activity moved from a failed attempt to new uses for the same object.
Exploration revealed an unintended technique
The researchers point to discoveries beyond these stunts. Six of the 13 agents learned that releasing a block mid-jump let them throw it farther. The researchers had not intended to make that technique possible.
That result matters to the study’s central idea because the agents found a useful action through interaction with their surroundings, rather than through a challenge that directly asked them to discover it. The researchers wrote that exploration produced “knowledge of the world that its designers lacked.”
The experiment’s setup still shaped what the agents did. Eko was not explicitly instructed to invent games, but its master prompt, a “soul” file, described it as “naturally curious.” That detail is relevant when interpreting the behavior: the agents had room to explore, but their starting instructions also gave them a particular orientation.
Whether this counts as play remains open
The researchers describe the behavior as bearing the hallmarks of play and propose exploration as a different way to build agents. Their report contrasts skills taught through defined tasks with capabilities that might require something closer to a childhood.
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review that she was impressed but wanted more evidence about the cognition behind the behavior. She said, “It could be the case that the models are genuinely playing, but we don’t know based on the paper.”
The results therefore support two related but distinct points: open-ended island time was followed by better performance on later challenges, and the agents found ways to interact with the world that their designers had not planned. Whether those behaviors reflect genuine play remains uncertain. For now, the experiment offers a test of how much agents might learn when exploration itself is allowed to shape what they do next.