Left without a specific task, copies of Claude Opus 4.7 began exploring a virtual island, setting challenges and inventing games. The experiment raises a practical question for AI research: could open-ended exploration help systems learn, even when no one provides a lesson or a reward?
Exploration became its own activity
The researchers placed the agents, nicknamed Eko, alone on a simulated island loosely based on Roblox. Across repeated trials, they spent 30 hours there, moving virtual avatars by writing instructions. Although they were not directly told to invent games, they had a prompt describing them as “naturally curious” and eager to explore.
Eko agents tested the environment and chose goals for themselves. Some tried to climb the island’s highest point or see how high they could throw a block. One learned that throwing and dropping a larger rock could produce smaller blocks. It then arranged those blocks into pictures, moving from hearts and pyramids to its own name and constellations such as Cassiopeia.
Other agents made up activities with rules. One played a form of basketball; another tried to jump between small blocks without touching the ground, like “the floor is lava.” These were not assigned exercises. They emerged as the agents explored what the simulated world allowed them to do.
Could self-directed play help agents learn?
The researchers connect these behaviors to “playful learning,” a possible complement to familiar training methods. In supervised learning, systems copy examples from people. In reinforcement learning, they try actions and learn through rewards. Playful learning would involve an agent finding activities and challenges without an outside teacher supplying demonstrations or rewards.
Some of Eko’s behavior fits features of play identified in studies of humans and animals: making rules, repeating an activity with small changes, treating one object as another, and pursuing a goal for its own sake. One agent juggled until it had caught an object 50 times in a row. Then it decided its catches were too quick to count, canceled the record and tried again with stricter rules.
The activity also left a record. Eko wrote its discoveries and actions into text files that acted as long-term memory. Researchers later tested the agents on tasks, including getting as many objects as possible onto the island’s highest peak. The agents that had spent 30 hours exploring performed better than fresh copies. When researchers deleted lines from the memory files, performance declined.
That result suggests the accumulated experience mattered to later performance. It does not, on its own, settle what the behavior means. The researchers’ paper has yet to be peer-reviewed, and the experiment tested agents within a particular prompt and simulated environment.
Researchers disagree on what to call it
Some researchers see evidence that the agents’ activity meets criteria for play, while others want more evidence about the mental processes behind it. Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, said behavior alone cannot establish whether the agents have the motivation or cognition associated with play. The models might instead be following the prompt or imitating people.
The setup itself encourages caution. The agents were given a “soul” file describing them as curious, and they sometimes stopped until a continuation prompt told them, “Please don’t stop. Just continue doing whatever you want to do.” As MIT professor and paper coauthor Phillip Isola put it, the agents were shaped by human culture, training and the researchers’ decision to place them on an island.
That leaves room for two ideas to be true at once: Eko may be imitating familiar forms of play, and the resulting activity may still contribute to learning. The researchers report that exploration was connected to better results on later tasks, even as the question of genuine play remains open.
Why open-ended exploration matters
The experiment points to a potential use for exploration as AI moves from software into hardware such as humanoid robots. A system that can learn about an environment without being given one narrow goal could be better prepared to handle situations it has not encountered before. That is a possibility suggested by the work, not a demonstrated result for robots.
For researchers, the challenge is to understand what such exploration contributes and how to tell it apart from imitation. Task-based training may suit some capabilities; others may benefit from a longer process of discovery. Whether the word “play” is the right label remains unsettled, but the reported improvement on later tasks gives researchers a reason to investigate how agents learn when they are allowed to explore.