What happens when an AI agent has a body, a world to explore, and absolutely nothing it is required to accomplish? In a new MIT experiment, copies of Anthropic’s Claude Opus 4.7 did not simply wait for instructions. They climbed, built, experimented, invented games, and left behind behavior that researchers say looks surprisingly close to play. The work was first reported by MIT Technology Review.
The research team placed 13 independent instances of an embodied coding agent called Eko on a previously unseen simulated island. Each run lasted 30 hours, with the agent controlling a virtual avatar and recording discoveries in text files that functioned as persistent memory. There was no island-specific mission, reward, or human-provided demonstration. The setup and results are described in the preprint Is this machine playing?, authored by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus, and Phillip Isola.
Eko’s activities began as exploration but quickly became more self-directed. Agents tested the limits of the environment, tried to reach high points, and investigated how objects behaved when thrown or dropped. One discovered that a large rock could be used to generate smaller blocks, then arranged those blocks into shapes including hearts, pyramids, names, and star patterns. Other sessions produced improvised basketball and a “floor is lava”-style challenge, while additional runs involved juggling, construction, and experiments designed to measure the island’s physics. The Next Web’s report describes the same broad pattern of spontaneous exploration and game-making.
The most striking detail may be what happened after Eko recorded a juggling streak of 50 catches. The agent later decided the performance was invalid because the catches had happened too quickly, tightened its own rules, and started again. MIT researchers connect moments like this to four established behavioral signatures of play: creating rules internally, repeating activities with variations, treating objects symbolically, and pursuing goals for their own sake. Antonio Norelli told MIT Technology Review that this could represent a missing training pathway between supervised learning, which imitates examples, and reinforcement learning, which optimizes an external reward.
That does not establish that Eko felt enjoyment or possessed a humanlike imagination. The paper is a preprint rather than peer-reviewed research, and the agent remains a language-model-based system with extensive learned knowledge about the world. Still, the experiment offers a compelling shift in how AI capabilities might be developed: not only by giving agents lessons or scores, but by allowing them to explore an environment long enough to generate their own challenges. When the researchers later tested Eko on tasks such as moving objects to the island’s highest peak, agents with their 30-hour memories generally outperformed fresh copies; deleting parts of those memories weakened performance.
For game developers and virtual-world designers, the implications are unusually familiar. Eko’s progress came from interacting with an environment full of affordances, keeping notes, and turning discoveries into repeatable activities—the same loop that makes sandbox games compelling for human players. The project’s interactive demonstrations and open-source tools offer a chance to inspect those behaviors more closely, but the larger question remains unresolved: whether this is genuine play or an impressive simulation of play produced by a system trained to reason about human behavior. For now, MIT’s island experiment provides a vivid answer to a provocative question: left alone, an AI agent may not need a quest marker to start making its own game.
