Researchers at MIT studied agents based on copies of the Claude Opus 4.7 language model, nicknamed Eko. In the experiment, agents were left alone for 30 hours on a simulated island loosely modeled on the game Roblox. The Eko copies could write instructions that moved their avatars around the island.
The paper describing the work has not yet been peer-reviewed.
Observed behaviors: goals, games and creations
The team reported several unexpected behaviors. Agents explored the boundaries of the simulated world and generated their own objectives: for example, trying to reach the map’s highest point or testing how high they could throw a block. One agent discovered that by throwing and dropping a larger rock it could create new blocks; using these, it began arranging patterns on the ground, starting with simple shapes like hearts and pyramids and later writing its own name and mapping constellations such as Cassiopeia. One Eko logged: “Tonight I built the night sky in stone.”
Another Eko iteration invented games including a basketball-like activity and a “the floor is lava”–style challenge where it had to jump among small blocks without touching the ground.
In one detailed example an Eko began juggling and recorded 50 consecutive catches. The agent then judged those catches to have happened too quickly to count as a valid record, deleted the record and restarted with stricter rules; the researchers interpret this behavior as analogous to human play, where participants set and then raise their own challenges.
Memory logs and impact on later performance
Agents logged activities and discoveries in text files that served as long-term memory. When researchers later tested them on specific tasks—such as getting as many objects as possible onto the island’s highest peak—agents that had spent 30 hours “playing” performed better than fresh copies without the same experience. The researchers also found that ablating (deleting) lines from an agent’s memory files degraded performance, suggesting the logs contributed causally to improved skill.
Framing as “playful learning” and critical perspectives
The authors argue Eko’s behavior meets four hallmarks associated with play in classic studies: inventing rules and challenges, repeating activities with slight variations, pretending objects were other objects, and pursuing goals for their own sake. They describe this mode as “playful learning,” a form of self-driv en exploration distinct from typical supervised or reinforcement-learning regimes.
However, several researchers urged caution. Gordon Burghardt, a psychologist at the University of Tennessee who has studied play in animals, agrees some criteria are met but remains cautious. Nathan Cloos (MIT), a coauthor, acknowledges that “all of this could be role-playing,” though the paper’s authors argue the demonstrated learning effects indicate the behavior is not merely hallucinated play.
Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, praised the work but said stronger evidence of underlying cognition and motivation would be needed to classify the behavior definitively as play. She noted that purely behavioral observations could reflect imitation of human-like playfulness rather than intrinsic motivation.
Experimental nudges and prompt design
The research setup intentionally biased agents toward human-like behaviors in several ways. Each Eko had a “soul” file—a text document that described the agent as “naturally curious” and eager to explore—which served as a master prompt shaping its disposition. Agents also sometimes remained inactive until the team provided a continuation prompt reading, “Please don’t stop. Just continue doing whatever you want to do.”
Phillip Isola (MIT), a coauthor, emphasized that the system did not spontaneously decide to play in a vacuum: it is a product trained on cultural data by Anthropic engineers and then placed on the island by researchers. Nonetheless, Isola argued, the surprising behaviors were causally related to the agents’ later skillful performance.
Why this matters
The authors suggest that as large language models move from software to embodied forms such as humanoid robots, exploration without a fixed external objective—akin to play—could help systems adapt flexibly to novel situations. The paper contrasts task-driven optimization with a developmental, play-like approach, noting: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”
Ultimately, the study’s primary claim is pragmatic: whether the agents are truly “playing” in the human sense may be less important than the extent to which their autonomous, playful-like exploration contributes to learning and competence.
Note on publication
The authors acknowledge the manuscript has not yet been peer-reviewed; further work is required to clarify the cognitive and motivational mechanisms behind the observed behaviors.
