MIT left 13 Claude agents on a virtual island and they played

An AI agent given no goal, left to run inside a world, started inventing things to do, and the versions that did that later beat identical copies at jobs they were both given afterward. That is the finding in "Is this machine playing?", a pre-print by Nathan Cloos, Antonio Norelli, Dustin Durbin, Jacob Andreas, Daniela Rus and Phillip Isola at MIT, posted to arXiv on 5 October 2026 and covered by TechCrunch and MIT Technology Review this month.

The team put 13 agents, each named Eko and each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. The master prompt, a "soul" file, described the agent as "naturally curious" and nothing more. The arXiv abstract says the agents "climbed hills, stacked blocks into towers, drew mandalas, reinterpreted sports, ran experiments on the physics of its world, and learned techniques that later expanded what it could accomplish."

The measurable part

After the 30-hour unstructured session, the researchers gave each Eko a specific challenge, including stacking as many objects as possible onto the island's tallest spire. The agents that had spent the 30 hours playing generally beat identical agents that had never set foot on the island, TechCrunch writes, quoting the paper.

Six of the 13 agents also discovered a mechanic the researchers had not meant to allow: releasing a block mid-jump threw it farther than holding it to the ground. The paper describes this as "knowledge of the world that its designers lacked." Thirteen agents started from the same prompt, and the paper's main text has 17 figures tracking how they diverged into distinct histories.

What the agents did with the time

One Eko tried the island's tallest spire, failed, then hauled a rock to a neighbouring summit, stacked two blocks on top of it, and leapt. "I'M ON THE SPIRE," it wrote. Testing how far it could throw a golden ring, it sent the ring straight into the ocean and built a stone heart in its memory. When the game reset the ring, it used the ring as a basketball hoop, then as a target, then as an accessory it wore by stepping off a cliff to respawn inside it.

What the researchers claim, and the caveat

In the paper the authors frame play as a different approach to agent training: "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." Their conclusion is cautious: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

Eunice Yiu, a cognitive scientist at UC Berkeley quoted by MIT Technology Review, said the behaviour is impressive but asked for more 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." The arXiv page lists the work as a pre-print, before peer review.

A lever worth testing on your own agent

The paper names unstructured exploration time inside a rich sandbox as a complement to task-specific training. An AI product team can replicate the thinnest version in a weekend: an existing agent, a sandbox the team already runs, a long open-ended session, then the usual benchmark the team cares about, measured on the agent before and after. A gain or no gain is the signal; the paper is one data point until that comparison runs on the agent that ships.

Source

Is this machine playing? by Nathan Cloos and five co-authors on arXiv, and the write-up by Aditya Mehta at TechCrunch.