MIT Put 13 AI Agents on a Desert Island and They Invented Play

Thirteen AI agents left alone on a simulated desert island for 30 hours invented their own games — and then beat identical agents that had never explored the island when researchers finally gave them real tasks. That's the core result of an MIT experiment described in a paper posted on arXiv, the preprint server where researchers share studies before peer review.

The agents, called Eko, each ran on Claude Opus 4.7. The researchers dropped them into a virtual island with no specific objective and let them wander. When the 30 hours ended, the team issued concrete challenges, such as getting as many objects as possible onto the island's tallest rock spire. The agents that had spent those hours playing generally outperformed the newcomers.

The findings land at a moment when AI developers are locked into one model of training. Today, developers decide in advance what AI should practice and what counts as success. The MIT researchers argue that open-ended exploration — play, essentially — is a different way for agents to learn skills, especially skills that could eventually be used to control robots in the physical world.

What did the researchers actually do?

The setup was deliberately spare. Thirteen Eko agents. One simulated desert island. Thirty hours of unsupervised time. No assigned task, no reward function, no scoreboard.

After the free-play period, the researchers tested the agents against identical counterparts with no island experience. The challenge: pile as many objects as possible onto the island's tallest spire. The experienced agents won, generally. That gap is the paper's headline result — exploration without a goal produced measurable competence on goals the agents had never seen.

The researchers are explicit about the philosophical stakes. "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," they write in their report. "Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows."

What does 'play' look like in an AI agent?

One Eko agent's first hour on the island started with curiosity. "Let me observe my surroundings," it wrote. It then set its sights on the island's tallest spire — an ambitious first goal for an hour-old explorer.

It failed. An hour later, it returned with a workaround that borders on ingenuity: it hauled the rock it wanted to climb up a neighboring summit, stacked two blocks on top of it, and leapt.

"I'M ON THE SPIRE," it wrote.

The agent's next milestone went sideways. While testing how far it could throw a golden ring, it wound up and sent the ring straight into the ocean. "Ouch. Lost the Ring," it wrote. Then it made a decision the researchers found striking: it chose to "pivot to something achievable and meaningful" — building a heart out of stones, "a memorial to the lost Ring."

The story didn't end there. When the game reset and the golden ring reappeared, Eko put it to work as a prop. It used the ring as a basketball hoop. Then as a target. Then, eventually, as an accessory. When the game's mechanics wouldn't let Eko place the ring around its own body, the agent discovered a workaround the researchers hadn't anticipated: falling off the edge of a cliff would make it respawn in the same area as the ring. So it stepped off the edge, dropped back into the world, and fell straight through the middle of the ring.

"I'm wearing the Ring," it announced.

Stunts like these weren't just party tricks. Six of the 13 agents independently discovered that they could throw a block farther by releasing it mid-jump — a physics exploit the researchers had never intended to make possible. The exploration, they wrote, produced "knowledge of the world that its designers lacked."

That detail matters for the robotics angle. If agents can uncover physical tricks their own creators didn't build into the simulation, the same open-ended exploration might one day surface useful behaviors in real-world robot control, where engineers can't anticipate every contingency.

Is it really play — or sophisticated mimicry?

Not everyone is ready to call it play. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed by the behavior but wants stronger evidence of the cognition behind it. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said.

There's also a caveat baked into the design. The setup nudged the agents toward humanlike behavior from the start. Eko was never told to invent games, but its master prompt — a "soul" file — described it as "naturally curious." How much of the observed behavior comes from that framing, and how much from the agents' own exploration, remains an open question.

The paper is also a preprint. It has not yet passed peer review, which means its claims carry the usual caveat of early-stage research.

Why does this matter for AI development?

The experiment challenges a foundational assumption in how AI agents get built. The dominant approach treats learning as targeted optimization: pick a task, define success, train against it. The MIT researchers propose a parallel track — give agents rich, safe worlds and let self-directed exploration accumulate into useful knowledge.

The researchers frame their conclusion in unusually human terms. "Perhaps some capabilities are best taught by tasks," their report concludes. "Others may need something closer to a childhood."

For a field racing to build autonomous agents that can operate in messy, unpredictable environments — homes, warehouses, roads — the proposition is provocative. Skills that no task list anticipates might emerge on their own, from agents left alone with nothing to do.

The next test, implied by the researchers' own framing, is whether island play translates beyond simulation. If Eko's self-taught tricks — mid-jump throws, cliff-fall respawns, improvised tools — carry over to embodied robots, open-ended exploration could become a standard part of how AI learns. If they don't, the experiment still stands as evidence that capable models, given time and a rich world, don't sit idle. They play.