In a simulation, stranded AI agents began to play

On a simulated island, a copy of Claude Opus 4.7 stood with a block in its hands and tested how high it could throw it. Nobody had asked it to. The agent, nicknamed Eko, was one of several placed alone on a virtual island for 30 hours, in an experiment that its designers compared to the films Cast Away and Home Alone. They had given the agents only the means to write instructions that moved their avatars.

What happened next is the surprising part. The agents set their own targets: scale the highest peak, or throw things as far as possible. One discovered that throwing and dropping a larger rock produced new blocks, and used them to build drawings: hearts, pyramids, then its own name and the constellation Cassiopeia. “Tonight I built the night sky in stone,” it wrote in its activity log. Another invented games, including a form of basketball. Eko also set a juggling record of 50 consecutive catches, then cancelled it because the catches had come too fast to count, and started again under stricter rules.

That last detail is what the MIT team, led by postdoc Antonio Norelli, wants to put forward. AI systems usually learn from human examples or from rewards. Here the learning came from the agents' own games. Their logs served as long-term memory, and the agents that had spent 30 hours playing outperformed fresh copies on tasks such as moving as many objects as possible to the island's peak. Deleting lines from their memory files degraded their performance. The team argues the play itself contributes to learning, and that this is the missing piece: a machine that learns without a human demonstrating or rewarding it.

The result is far from settled. The paper has not yet been peer-reviewed. Psychologist Gordon Burghardt, whose work on play the team relies on, agrees that the behavior meets some of his criteria, but asks whether the model was simply role-playing what a playful agent would do. Nathan Cloos, an MIT author, concedes that “all of this could be role-playing.” The setup also nudged the agents: they had a “soul” file describing them as “naturally curious,” and they often stopped until a prompt told them to keep going. Eunice Yiu of UC Berkeley, who studies exploration, says that calling it play would require more evidence of motivation behind it.

For now, the practical stake is the next generation of AI. As language models move into hardware such as humanoid robots, the team suggests that exploring without a fixed goal could help them adapt to unexpected situations. That benefit is a hypothesis, not a result: the experiment ran entirely in simulation, and no robot has yet been trained this way. Researchers working on world models, which are internal representations of how environments behave, may be the first to test it. The paper's closing line proposes that some skills are best taught by tasks, while others may need something closer to a childhood.