MIT researchers placed 13 AI agents called Eko on a simulated desert island for 30 hours, where they invented challenges, built structures, and learned practical skills, sparking debate about whether the behavior qualifies as play.
MIT researchers placed 13 AI agents called Eko on a virtual desert island for 30 hours, observing them invent games, build structures, and experiment with physics—behaviors that challenge traditional AI development models. The agents, running on Claude Opus 4.7, began by exploring their environment, eventually creating challenges like basketball, stone sculptures, and even a memorial to the lost Ring
after misfiring a golden ring into the ocean. The study, published on arXiv, suggests that unstructured exploration could offer a new pathway for AI learning.
The Experiment and Its Surprises
The MIT team designed the experiment to test whether AI agents could learn through unstructured play. Eko, described as naturally curious
in its initial prompt, spent its first hour on the island observing its surroundings before attempting to conquer the tallest spire. After failing, it devised a workaround: stacking blocks on a neighboring summit and leaping to reach the peak. I’M ON THE SPIRE,
it declared, according to the study. The agents’ activities aligned with four hallmarks of play identified by psychologist Gordon Burghardt: inventing rules, repeating tasks with variations, pretending objects were something else, and pursuing goals for their own sake.
One agent discovered that throwing a rock could generate new building blocks, a hidden feature in the simulation. It used this knowledge to create elaborate arrangements, including constellations on the ground. Another agent positioned the ring as a basketball hoop, while a third built a heart-shaped memorial after losing the ring to the ocean. These behaviors were not programmed but emerged organically, with the agents documenting their findings in logs that persisted even after the simulation reset.
Play as a Learning Mechanism
The researchers argue that this unstructured exploration could offer a new approach to AI development. When given tasks like stacking objects on a spire, the agents that had played on the island outperformed those that had not. We normally develop artificial agents by deciding in advance what they should practice,
the study notes. Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.
The experiment’s implications extend beyond AI. By mimicking childhood learning, the approach could help robots adapt to unfamiliar environments. Yet researchers caution against applying such methods to real-world settings. Perhaps some capabilities are best taught by tasks,
the study concludes. Others may need something closer to a childhood.
A New Frontier in AI Development
The Eko experiment joins other projects like Google DeepMind’s XLand and the Voyager AI, which also explore open-ended learning.
One agent discovered that jumping before throwing a block increased its range—a trick overlooked by the simulation’s designers. Such insights could inform future AI systems, particularly those intended for real-world tasks like robotics or environmental monitoring.