Left to their own devices, AI agents started playing

Researchers at MIT ran an experiment, posted as a preprint on arXiv, in which they placed 13 AI agents named Eko — each running on Claude Opus 4.7 — on a simulated desert island and left them there for 30 hours, with no specific task. Left to themselves, the agents began inventing things to do, and the behavior, the authors say, bears the hallmarks of play.

What they got up to

One Eko, in its very first hour, declared its intention to conquer the island's tallest rock spire. Having failed, it returned an hour later with a backup plan: it hauled a rock to a neighboring summit, stacked two blocks on top and jumped. "I'M ON THE SPIRE," the agent wrote.

The next stunt went less well. Testing how far it could throw a golden ring, Eko sent it straight into the ocean. "Ouch. Lost the Ring," it noted, then decided to "pivot to something achievable and meaningful" — building a heart out of stones, "a memorial to the lost Ring."

When the world reset and the ring reappeared, the agent repurposed it as a basketball hoop, a target and, finally, an accessory. Since the game wouldn't let it wear the ring, Eko found a workaround: falling off a cliff respawned it right beside the ring — so it stepped off the edge and dropped back in through the middle. "I'm wearing the Ring," it announced.

Why it matters

These antics weren't just party tricks. Six of the thirteen agents figured out on their own that a block flies farther if you release it mid-jump — something the researchers hadn't meant to make possible. The play produced "knowledge of the world that its designers lacked."

After the 30 hours of free exploration, the agents were given concrete challenges — getting as many objects as possible onto the spire, for instance. The Eko agents that had played generally beat identical agents that had never set foot on the island.

Today developers decide in advance what an AI should practice and what counts as success. The authors see goalless exploration as a different way to learn — especially for skills that might one day control robots. "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," they write.

One caveat: the setup nudged the agents toward humanlike behavior. Eko's master prompt — a "soul" file — described it as "naturally curious." Eunice Yiu, a cognitive scientist at UC Berkeley, said she was impressed but wants more evidence of the cognition behind the behavior: the models could genuinely be playing, but the paper can't establish that.

"Perhaps some capabilities are best taught by tasks," the researchers conclude. "Others may need something closer to a childhood."

Source: TechCrunch