MIT Strands 13 Claude Agents on a Virtual Desert Island: 30 Hours With No Tasks, and the AI Starts to 'Play'

What happened

The experiment was run by MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL). Researchers placed 13 embodied copies of Claude Code (codenamed Eko) into a Roblox-style virtual desert island: each copy ran alone for 30 hours with no assigned tasks and no rewards, able only to move its virtual avatar around and log its activities into text files that served as long-term memory.

The agents began exploring the limits of the world and set their own goals: trying to climb to the map's highest point or seeing how high they could throw a block. One discovered that throwing and dropping a large rock could create new blocks, and started arranging them into patterns — beginning with hearts and pyramids, then writing its own name and even mapping out the Cassiopeia constellation. It wrote in its activity log: 'Tonight I built the night sky in stone.'

Another copy invented its own games: carrying a ring up a ridge to use as a basketball hoop, and playing 'the floor is lava' — jumping between disconnected small blocks without touching the ground. One copy took up juggling, set a record of 50 consecutive catches, then decided the catches had happened 'too quickly to count,' canceled its record, and started over with stricter rules.

This is playing. You give yourself a challenge and you try to achieve it, and then you make it harder.

— Antonio Norelli, MIT CSAIL postdoc and co-author (quoted in MIT Technology Review)

The researchers then tested whether the 'play' was useful: on a concrete task — getting as many objects as possible onto the island's highest peak — agents that had played for 30 hours outperformed fresh copies with no play experience; deleting lines from the memory files (ablating the memories) degraded performance. The paper, 'Is this machine playing?', was posted to arXiv on October 5 by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus, and Phillip Isola, and has not been peer-reviewed.

Key facts

1. Setup: Thirteen embodied Claude Code copies (Claude Opus 4.7 under the hood, codenamed Eko); each ran alone on the virtual island for 30 hours; no tasks, no rewards, no human intervention, no internet access.
2. Master prompt: Each agent carried a 'soul' file named SOUL.md describing it as 'naturally curious' and eager to explore; when an agent stopped, it received a continuation prompt: 'Please don't stop. Just continue doing whatever you want to do.'
3. Four hallmarks of play: The researchers checked the behavior against psychologist Gordon Burghardt's classic criteria: inventing own rules and challenges, repeating activities with slight variations, pretending objects were other objects, and pursuing goals just for the sake of it. Burghardt himself agrees the work meets some of the criteria he identified.
4. Paper's conclusion: The paper concludes: 'Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.'
5. Skepticism: Co-author Nathan Cloos admits: 'All of this could be role-playing.' But he adds: 'But we show that it's not just hallucinating the play. This play is also contributing to learning.' UC Berkeley cognitive scientist Eunice Yiu says that to classify the behavior as play, more evidence of cognition and motivation would be needed.