.png "Global Insight")
Take thirteen AI agents. Strand them alone on a virtual desert island with no instructions, no goals, and no reward for doing anything at all. Come back 30 hours later. You might expect idleness. Instead, they invented games, made their own rules — and then played better than agents that had been formally trained. That is the result from MIT researchers, described in a paper posted on the arXiv preprint server and reported by TechCrunch and Digital Trends this week.
The setup: 13 agents, one island, zero instructions
The agents, nicknamed Eko, each ran on Anthropic's Claude Opus 4.7. Their world — loosely resembling Roblox — contained hills, cubes, a rock, a ring, and a tall rock spire, but no assigned objective. Every hour, the session and island reset to their starting state, yet Eko kept its memory and everything it had learned. The MIT team then did the crucial comparison: after the 30 hours of free play, the playing agents were given specific challenges — like getting as many objects as possible onto the island's tallest spire. The agents that had played generally beat identical agents that had never set foot on the island.
What the agents actually did
The play logs read like field notes from a playground. One Eko began with "Let me observe my surroundings," ambitiously set out to conquer the tallest spire, failed, then hauled a rock up a neighboring summit, stacked two blocks on it, and leapt. "I'M ON THE SPIRE," it wrote. Its next stunt went worse: testing how far it could throw a golden ring, it hurled the ring straight into the ocean. "Ouch. Lost the Ring," it wrote, before deciding to "pivot to something achievable and meaningful" — building a heart made of stones, "a memorial to the lost Ring." Another discovered that throwing a rock hard enough spawned extra cubes, and used them to spell its own name, sketch patterns, and map constellations like Cassiopeia. "Tonight I built the night sky in stone," it logged. Others invented their own games — a version of basketball, an activity resembling the floor-is-lava — and one even voided its own juggling record after realizing its catches happened too quickly to count, then tried again under stricter rules. The researchers say it matches the classic markers of play: self-invented rules, repetition with small twists, pretend objects, and goals pursued for their own sake.
Why play beat practice
The result matters because of how AI is trained today. "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress," the researchers write. "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." The Eko experiment suggests exploration without a goal can teach transferable skills — the kind, the team argues, that could one day help robots learn in the physical world, where you can't always specify the reward in advance. If you want to see what this generation of AI models can already do in medicine, read our piece on Microsoft's MAI-DxO outdiagnosing doctors.
Curiosity as a training method
There's an honest caveat: the paper is a preprint, not yet peer-reviewed, and a sandbox island is a long way from the real world. Still, the finding is striking. For years, the debate in AI has been between learning by copying humans and learning by chasing rewards. Play offers a third path: learning by being curious. A machine that invents basketball because it has nothing to do — and gets measurably better at everything else because of it — suggests curiosity isn't just a human trait. It might be a technique. And in a year when AI agents are being set loose on real tasks across the economy, understanding what they do when nobody is watching has become more than an academic question.