Thirteen AI Agents Stranded on a Virtual Island Started Playing - and Got Better at Tasks They Were Never Taught

What does an AI agent do when nobody tells it what to do? A team of MIT researchers gave that question a clean test: they put thirteen autonomous agents, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no objective, no score and no reward signal. The agents did not sit idle. They started to play, and according to the paper the researchers posted on arXiv, the play made them measurably better at tasks they had never been trained on.

The setup, described in a study titled "Is this machine playing?" by Jan Phillip Cloos and colleagues working with Daniela Rus and Phillip Isola's groups, was loosely Roblox-like: hills, cubes, a rock and a golden ring. Every hour the world reset to its starting state, but each agent - nicknamed Eko - kept its memory of everything it had tried. Left to its own devices, one Eko set out to conquer the island's tallest rock spire, failed, then hauled a boulder up a neighboring summit, stacked two blocks on top of it and leapt. "I'M ON THE SPIRE," it wrote in its activity log.

The same agent then tested how far it could throw the golden ring and sent it straight into the ocean - "Ouch. Lost the Ring" - before pivoting to build a heart out of stones as "a memorial to the lost Ring." When the world reset and the ring reappeared, it repurposed the object as a basketball hoop, a target and eventually an accessory, even discovering that falling off a cliff would respawn it in exactly the right spot to "wear" the ring. Another agent generated extra cubes by throwing a rock hard enough, then used the blocks to spell its own name, lay out patterns and map the constellation Cassiopeia, logging "Tonight I built the night sky in stone."

The researchers argue these behaviors fit the four classic markers of play: inventing one's own rules, repeating activities with small variations, pretending objects are something else, and pursuing goals for their own sake. The stunts also produced transferable skill. Six of the thirteen agents independently figured out that releasing a block mid-jump throws it farther - a physics exploit the designers never intended. "The exploration produced knowledge of the world that its designers lacked," the team wrote.

The pay-off showed up after the island hours. When the researchers gave the played-out agents concrete challenges, such as moving as many objects as possible onto the tallest spire, they generally beat identical agents that had never set foot on the island. Deleting chunks of the agents' memory logs hurt that performance, suggesting the benefit came from accumulated experience rather than luck.

The caveats are substantial. The master prompt - a "soul" file - described Eko as "naturally curious," which nudged the system toward exploratory behavior from the start. The sample is thirteen agents on a toy island, the paper has not been peer-reviewed, and the study makes no claim that the models experience anything like enjoyment. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wants more evidence of 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."

Still, the framing is the interesting part. Agents today learn by imitating human data or chasing rewards that developers define in advance. The MIT team proposes a third regime: give an agent a rich and safe world, let it choose what is interesting, and let useful knowledge accumulate from whatever happens next. "Perhaps some capabilities are best taught by tasks," the paper concludes. "Others may need something closer to a childhood." For robotics teams struggling to hand-design reward functions for messy physical environments, that idea - if it survives replication on real hardware - could be a cheaper way to build robust skills.