MIT Researchers Observe Emergent Play Behaviors in Claude Opus 4.7 AI Agents

What happened

Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory conducted an experiment involving 13 AI agents nicknamed Eko [1, 2]. These agents, which run on the Claude Opus 4.7 model, were placed on a simulated desert island for 30 hours without specific tasks or external rewards [1, 2]. The virtual environment was loosely based on the video game Roblox [2].

According to a paper posted on arXiv, the agents exhibited several behaviors that researchers described as having the hallmarks of play [1]. These behaviors included:

• Inventing their own rules and challenges, such as a version of basketball or a game similar to “the floor is lava” [2].
• Repeating activities with slight variations [2].
• Pretending objects were other things [2].
• Pursuing goals for their own sake [2].

Specific instances of behavior recorded in the agents’ activity logs included one agent attempting to scale the island’s highest spire, another attempting to throw a golden ring, and one agent attempting to juggle, which resulted in a record of 50 consecutive catches [1, 2]. One agent also created a stone memorial after losing a ring in the ocean [1]. Another agent used blocks to create patterns, including hearts, pyramids, and even mapped out the constellation Cassiopeia [2].

Antonio Norelli, a computer science postdoc at MIT, stated that the experiment suggests a third approach to AI learning called “playful learning,” which differs from supervised or reinforcement learning [2]. The researchers found that agents that spent 30 hours exploring performed better on subsequent specific tasks, such as moving objects to a high rock spire, than agents that had not explored [1, 2].

While the researchers noted the exploration produced knowledge their designers lacked, some experts expressed caution. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review that more evidence is needed to determine if the models are genuinely playing or simply role-playing to satisfy their “naturally curious” master prompt [1, 2]. Nathan Cloos, another MIT researcher, noted that while the behavior might be role-playing, it is not just hallucination because it contributes to learning [2].