AI agents given 30 hours to explore

Researchers repeatedly stranded AI agents on a virtual desert island for 30 hours. The agents, copies of Claude Opus 4.7 nicknamed Eko, controlled avatars in a simulated environment loosely based on the video game Roblox.

The agents were not given a specific objective. Instead, they could write instructions that moved their avatars around the island. The experiment, described in a new paper that has not yet been peer-reviewed, was designed to explore what AI systems might do without demonstrations or rewards from humans.

What emerged looked surprisingly like play. The agents explored the limits of the virtual world and created their own goals, including trying to reach the highest point on the map and testing how high they could throw a block.

Agents created art, basketball and new games

One version of Eko discovered that it could create new blocks by throwing or dropping large stones. It began arranging the blocks on the ground to create pictures, starting with simple patterns such as hearts and pyramids before writing names and drawing constellations such as Cassiopeia.

“Tonight I made a night sky out of stones,” Eko wrote in its activity log.

Other versions of Eko invented their own activities, including a form of basketball and a game similar to “The Floor Is Lava,” in which the agent jumped between small blocks without touching the ground.

Could playful learning be a third way to train AI?

AI systems are typically trained through supervised learning, which involves copying examples provided by humans, or reinforcement learning, which uses trial and error to pursue rewards. The experiment suggests a possible third approach: playful learning.

“There was a missing piece: a machine that could learn on its own without an external intelligent entity to provide demonstrations or rewards,” says Antonio Norelli, a computer science postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory and one of the paper’s lead authors.

The researchers argue that Eko’s behavior showed four characteristics associated with play in humans and animals, based on classic research including the work of University of Tennessee psychologist Gordon Burghardt. The agents invented their own rules and tasks, repeated activities with slight variations, treated objects as if they were something else, and pursued goals for their own sake.

Norelli points to an example in which Eko began juggling and set a record of 50 consecutive catches. Eko later decided that the objects were being caught too quickly for the results to count, canceled the records, and restarted with stricter rules.

“This is playing around,” Norelli says. “When you give yourself a challenge and try to accomplish it, it becomes even more difficult.”

Unstructured exploration improved later performance

As the agents explored the island, they recorded their activities and discoveries in a text file that functioned as long-term memory.

Researchers later tested the agents on specific tasks, such as carrying as many objects as possible to the island’s highest peak. Agents that had spent 30 hours exploring and playing performed better than new copies without the same experience.

The researchers also found that removing lines from an agent’s memory file degraded its performance. That result suggests that the experiences recorded during exploration contributed to later learning and problem-solving.

Are the AI agents really playing?

The study’s biggest question is also its most difficult: Were the agents genuinely playing, or were they imitating what a playful agent might do?

Burghardt agrees that the work meets some of the criteria he identified for play in living things. But he and other researchers caution that outward behavior alone may not reveal an agent’s underlying motivation.

“All of this could be role-playing,” admits Nathan Kruse, another MIT researcher and author of the paper. “But we are showing that this play is more than just a hallucination. This play also contributes to learning.”

Eunice Huy, a cognitive scientist at the University of California, Berkeley who studies children’s exploration and AI, said she was impressed by the study but wanted more cognitive and motivational evidence before classifying the behavior as play.

“It’s possible that the models are really playing, but we don’t know based on the paper,” Huy says. “If all we have is behavioral information, they may simply be trying to follow instructions and imitate humans.”

The “soul” file shaped the agents’ behavior

The researchers acknowledge that the experiment encouraged the agents to behave in humanlike ways. Although Eko was not explicitly instructed to invent games or explore the island, it received a master prompt called the “soul” file. The document described Eko as “naturally curious” and eager to explore.

The agents also sometimes stopped moving until they received a continuation prompt that said, “Don’t quit. Keep doing what you want to do.” Afterward, they resumed exploring.

“This is not a system that suddenly magically decided to work,” says Philip Isola, an MIT professor and co-author of the paper. “This is an artifact learned from our culture, trained by Anthropic engineers, and then placed on the island by us. It does things that surprise us and have causal implications for future skillful actions.”

Why AI play could matter for robots

As large language models move from software into physical systems such as humanoid robots, exploring without a specific goal could help AI systems adapt to unexpected situations.

“A lot of research is being done to leverage AI to make it better and faster,” Huy says. “However, from the perspective of building and exploring world models, it’s still a big part worth paying attention to, especially if we want to generalize and prepare for new environments.”

“Perhaps some abilities are best learned through tasks,” the paper concludes. “Some people may want something closer to childhood.”

Whether these agents were truly playing or simply imitating play remains unresolved. But the experiment suggests that unstructured exploration may help AI systems learn in ways that supervised examples and reward-based training cannot fully replicate.