Autonomous AI agents left without tasks in a virtual world began inventing goals, building patterns and creating games, behavior researchers say resembles play and could aid “playful learning.”
Left alone with no explicit assignment, a group of AI agents did not sit idle. Instead, they began to explore, invent challenges, and even create games—behavior the researchers behind a new study argue resembles play.
In the experiment, scientists repeatedly placed autonomous agents—copies of Anthropic’s Claude Opus 4.7 that the team nicknamed “Eko”—onto a sparse virtual “desert island” for 30-hour stretches. The setup, which the researchers likened to cinematic isolation scenarios such as Cast Away or Home Alone, used a simulated environment loosely inspired by the game Roblox. Eko could write instructions to control an avatar and manipulate objects in the world.
What emerged surprised the team. Rather than waiting for direction, the agents started probing what was possible on the island and set their own objectives. Some attempted to reach the highest point on the map. Others tested physics-like constraints—for instance, trying to determine the maximum height they could launch an object.
One particularly striking behavior appeared after an agent realized it could generate new blocks by throwing and dropping a larger rock. With that “mechanic” uncovered, the agent began arranging blocks into ground patterns: first simple shapes such as hearts and pyramids, then more elaborate creations. Over time it spelled out its own name and even laid out star patterns resembling constellations, including Cassiopeia. In its log, the agent wrote: “Tonight I built the night sky in stone.”
Other versions of Eko improvised structured games. One created a basketball-like activity. Another devised an obstacle routine reminiscent of “the floor is lava,” requiring the agent to hop among small blocks without touching the ground.
The researchers see these behaviors as pointing toward a potential training paradigm beyond the two dominant approaches used today. Most AI systems learn either through supervised learning—imitating examples provided by humans—or through reinforcement learning—optimizing behavior by trial and error against an external reward signal. The new work suggests what the team calls “playful learning,” where an agent generates its own goals and constraints without demonstrations or rewards provided by a separate intelligent overseer.
“There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards,” said Antonio Norelli, an MIT Computer Science and Artificial Intelligence Laboratory postdoctoral researcher and one of the paper’s lead authors. The study has been released as a paper but has not yet undergone peer review.
To support the “play” framing, the team points to classic criteria proposed by psychologist Gordon Burghardt of the University of Tennessee. The researchers argue Eko exhibited four common features associated with play in animals and humans: self-imposed rules and challenges, repetition with variation, treating objects as if they were something else, and pursuing activities for their own sake. Norelli points to one instance in which Eko practiced “juggling” and reached 50 catches—then decided the attempt was too fast to be fair, voided the record, and restarted with stricter conditions.
Throughout each 30-hour session, the agents wrote text logs that functioned as long-term memory. When the team later evaluated the agents on concrete tasks—such as moving as many objects as possible to the island’s highest peak—those that had spent time freely exploring performed better than newly instantiated agents without that experience. Performance also worsened when researchers “ablated” memory by deleting lines from the stored logs.
Still, the findings raise a thorny question: is this genuine play, or sophisticated role-play? Burghardt said the results align with some of his play criteria, but MIT researcher Nathan Cloos, also an author, acknowledged the ambiguity. “All of this could be role-playing,” he said—while arguing the key result is that the behaviors appear to improve learning outcomes, not merely generate playful narration.
Eunice Yiu, a UC Berkeley cognitive scientist who studies exploration in children and AI, called the work impressive but urged caution in labeling it play without stronger evidence for underlying motivation and cognition. The agents’ behavior could reflect genuine play, she noted, but the field still lacks definitive proof.
Why it matters
If unstructured “play” can reliably build skills and transferable knowledge, it could become a new ingredient in training AI agents that must operate in open-ended environments. The research also highlights how memory—here, simple text logs—can shape competence over time, and how fragile that competence may be if memories are removed. Next questions include whether playful learning scales to more complex worlds and whether it leads to safer, more predictable behavior—or the opposite.