MIT researchers observe autonomous AI agents engaging in play

Researchers at the Massachusetts Institute of Technology (MIT) conducted an experiment titled “Is this machine playing?”, observing how AI agents behave in virtual environments without specific instructions or rewards. Using thirteen independent instances of an agent named Eko, powered by Anthropic’s Claude Opus 4.7, the team placed each agent on a simulated virtual island for 30 hours.

Despite identical starting conditions, the agents developed diverse, spontaneous behaviors. Some agents engaged in activities resembling sports, such as bowling or basketball, using available objects like blocks and rings. Others demonstrated more complex behaviors, including building structures like monuments, temples, or even spelling out their names. One agent even interpreted a system reset as the “end of a pilgrimage” after falling into the sea.

The study suggests these activities mirror biological play, where agents explore physical rules and develop hypotheses through trial and error. Notably, the researchers found that agents that engaged in this undirected play performed better on subsequent structured tasks, such as stacking objects on high terrain, compared to those that did not play.