Left to Themselves on a Virtual Island, AI Agents Invented Their Own Games

Give AI agents a simulated desert island, no task, and 30 hours - and they start making up their own challenges. MIT researchers say the behavior looks a lot like play.

The team placed 13 agents, each named Eko and running on Claude Opus 4.7, on the island and described the work in a paper posted to arXiv, which hosts studies before peer review. Later, when given specific goals - such as getting as many objects as possible onto the island's tallest rock spire - the agents that had spent those 30 hours playing generally outperformed identical agents that had never been on the island.

One Eko began by announcing it would observe its surroundings, then set out to conquer the tallest spire. An hour later, having failed, it hauled a rock up a neighboring summit, stacked two blocks on it, and leapt. "I'M ON THE SPIRE," it wrote. A later experiment with a golden ring ended in the ocean - "Ouch. Lost the Ring." - so the agent built a stone heart as "a memorial to the lost Ring." When the game reset and the ring returned, Eko used it as a basketball hoop, a target and, finally, an accessory, stepping off a cliff so it would respawn through the middle of the ring: "I'm wearing the Ring."

Six of the 13 agents also discovered they could throw a block farther by releasing it mid-jump - something the researchers had not intended to make possible. The exploration produced "knowledge of the world that its designers lacked," they wrote.

The setup nudged the agents toward humanlike behavior: Eko was never told to invent games, but its master prompt described it as "naturally curious." Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior, saying it could be that the models are genuinely playing, but the paper doesn't establish it.

The researchers see goal-free exploration as a different way for agents to learn skills, including ones that could one day be used to control robots. "Perhaps some capabilities are best taught by tasks," their report concludes. "Others may need something closer to a childhood."