MIT researchers repeatedly left copies of Claude Opus 4.7 alone on a simulated desert island for 30 hours at a time, MIT Technology Review reports. Nicknamed Eko, the agents could write instructions to move their virtual avatars around the island. One iteration devised a version of basketball and a game resembling “the floor is lava,” jumping between blocks without touching the ground.
The agents were not explicitly told to invent games or explore. But the setup was not prompt-free: a master “soul” file described Eko as “naturally curious” and eager to explore. The agents also frequently stood still until a continuation prompt urged them to keep doing whatever they wanted.
What carried over to later tasks
As the agents explored, they logged activities and discoveries in text files that served as long-term memory. On later specific tasks, including getting objects onto the island’s highest peak, agents with 30 hours of island experience performed better than fresh copies without that experience, Technology Review reports. Deleting lines from their memory files degraded performance.
Was it play?
The researchers argue that Eko showed hallmarks associated with play, including inventing challenges and repeating activities with variations. Antonio Norelli, an MIT Computer Science and Artificial Intelligence Laboratory postdoc and one of the paper’s lead authors, points to an agent that canceled its own juggling record after deciding its catches had happened too quickly to count, Technology Review reports.
Nathan Cloos, another MIT researcher and paper author, acknowledges that “All of this could be role-playing,” while arguing that the activity contributed to learning. UC Berkeley cognitive scientist Eunice Yiu says the behavior alone does not establish the cognition and motivation needed to classify it as play. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” she says. The paper has yet to be peer-reviewed.
