MIT researchers stranded Claude agents on a virtual island and they invented games

Agents with nothing to do started playing

What does an AI agent do when nobody assigns it a task? According to MIT Technology Review, researchers at MIT repeatedly left AI agents alone in a virtual desert-island simulation for 30 hours at a time, and the agents responded by setting their own goals, building designs out of blocks and inventing games.

The subjects were copies of Anthropic's Claude Opus 4.7, nicknamed Eko, which steered avatars around an island environment loosely modeled on Roblox. With no explicit instructions, they probed the edges of their world, tried to reach the map's highest point and tested how high they could throw objects.

One agent discovered that throwing and dropping a large rock generated new blocks, then used the trick to lay out drawings on the ground, starting with hearts and pyramids and progressing to its own name and constellations such as Cassiopeia. "Tonight I built the night sky in stone," it wrote in its activity log, per MIT Technology Review. Another instance of Eko devised games, including a basketball variant and one resembling "the floor is lava," hopping between small blocks without touching the ground.

A possible third path to learning

Most modern AI is trained by supervised learning, copying human-provided examples, or reinforcement learning, pursuing a reward through trial and error. MIT Technology Review reports that the experiment points toward a third approach the researchers call playful learning.

"There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards," Antonio Norelli, a postdoc at MIT's Computer Science and Artificial Intelligence Laboratory and lead author on the paper, told MIT Technology Review. The paper has not yet been peer-reviewed.

The team measured Eko's behavior against classic play research, notably the framework of University of Tennessee psychologist Gordon Burghardt, and found it displayed the four hallmarks of play seen in humans and animals: it invented its own rules and challenges, repeated activities with slight variations, treated objects as stand-ins for other objects, and pursued goals for their own sake. Norelli cites a juggling episode in which Eko logged a record of 50 consecutive catches, decided they had happened too quickly to count, voided the record and started over under stricter rules. "You give yourself a challenge and you try to achieve it, and then you make it harder," he says.

The play produced measurable skill

As the agents explored, they recorded activities and discoveries in text files that served as long-term memory. When researchers later tested them on concrete objectives, such as moving as many objects as possible onto the island's highest peak, the agents that had spent 30 hours playing outperformed fresh copies of the model. Deleting lines from the memory files, a process the researchers call ablation, degraded performance, suggesting the play experience itself, preserved in memory, drove the gains.

Or is it role-playing?

Whether any of this counts as genuine play is contested. Burghardt agrees the work meets some of his criteria for play in living creatures, but the open question is whether a language model is truly playing or merely imitating what a playful agent would do in this situation.

"All of this could be role-playing," coauthor Nathan Cloos admits, "but we show that it's not just hallucinating the play. This play is also contributing to learning."

Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, told MIT Technology Review that the findings are impressive but that calling the behavior play would require evidence of cognition and motivation underneath it. With behavioral data alone, the agents could simply be following the prompt or imitating humans, she says. The setup did push the agents toward human-like conduct: a master "soul" file described Eko as naturally curious and eager to explore, and the agents often stood idle until a continuation prompt told them to keep going.

Coauthor Phillip Isola, an MIT professor, frames it plainly: this is not a system that suddenly decided to play. It is an artifact trained on human culture by Anthropic's engineers and then placed on an island by researchers. Even so, he says, "it does things that surprise us and that are causally related to its future skillful behavior."

Why it matters

As large language models move from software into hardware such as humanoid robots, the ability to explore a space without a specific goal could help them adapt flexibly to unexpected situations. Yiu notes that most current work aims at exploiting AI, making it faster and better, while exploration and world-model building remain underfunded attention despite being central to generalizing to new environments. The paper's own conclusion captures the stakes: some capabilities are best taught through tasks, while others may need something closer to a childhood.