Alone on a virtual island, Claude stopped working and started playing.

MIT built a tiny virtual island — hills, some blocks, a rock, a ring — and dropped in 13 copies of a Claude agent called Eko. No instructions. No score to chase. No reward. Just thirty hours alone on the island, each. So it started playing. Eko poked at the limits of its little world like a kid in a backyard. It learned that throwing a rock hard enough spawned new cubes, then used them to spell its own name and map constellations — Cassiopeia, in stone. Its activity log reads: "Tonight I built the night sky in stone." It invented basketball. It played the floor is lava. It set a juggling record of 50 catches, then tore the record up because the catches came too quickly to really count — and started over with stricter rules. Researchers say the behavior fits the classic hallmarks of play: made-up rules, repetition with variations, pretend, and chasing goals just because. But here's the part that matters: when they later tested the agents on actual tasks — like hauling objects to the island's highest peak — the ones that had played beat the fresh copies. Deleting lines from their memory logs made them worse. The team calls it "playful learning": a third way besides copying humans and chasing rewards. Their closing line: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood." The honest caveat, and it's the researchers' own: Eko had a "soul" file calling it naturally curious, and sometimes it just stood there until someone nudged it with "Please don't stop." It might be role-playing. But the play wasn't fake — it made the agent better. And that distinction might matter more than the question of whether it "really" played. We're building minds, then arguing about whether they're allowed to have fun. The evidence says: let them play. It works. Sources: MIT Technology Review (Amit Katwala), "AI researchers stranded Claude on a virtual desert island. It started to 'play.'", Oct 19, 2026; Digital Trends (Manisha Priyadarshini), "What does an AI agent do when it's left alone with no task? This experiment offers a clue", Oct 19, 2026; arXiv preprint "Is this machine playing?" (Cloos, Norelli, Durbin, Andreas, Rus, Isola, MIT CSAIL), posted Oct 5, 2026.