Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) left 13 independent copies of an agent called Eko — based on Claude Code running Anthropic’s Claude Opus 4.7 — alone on a virtual island for 30 hours each, with no task, reward or defined objective.
What happened?
The environment simulates an island with rigid-body physics, cubes, a rock, a ring and two elevated points (Peak and Spire). Agents observed the world only through textual JSON descriptions and acted with commands such as move, jump, pick up, place and throw. Without explicit instructions to explore or play, they developed spontaneous behaviors: they climbed the highest point, stacked blocks into towers and monuments, drew patterns (including hearts, pyramids and even their own name "EKO" in stones), mapped constellations and invented games resembling basketball and "the floor is lava".
One agent logged 50 rounds of juggling and then redefined the rules because it considered the pace "too fast to really count". Others discovered that throwing the rock hard generated new cubes — a mechanic the researchers had built into the simulator that the agents found on their own. Some also found that throwing a block mid-jump made it fly much higher.
Why it matters
The paper, titled "Is this machine playing?" and posted to arXiv (2610.07130) on October 5, 2026 (not yet peer-reviewed), argues that the behaviors meet classic criteria for play described by psychologist Gordon Burghardt: inventing rules, repeating with variations, pretending and pursuing goals for their own sake. More importantly, the experience generated transferable learning. In later tests (such as placing as many objects as possible on the Spire), agents that spent the 30 hours performed better than fresh copies of the same model — an average of 10.3 objects versus 3.5. Deleting specific lines from the agents’ textual memory degraded performance on the tasks that depended on that discovery.
The researchers, led by Nathan Cloos and Antonio Norelli (with participation from Jacob Andreas, Daniela Rus and Phillip Isola), note that this points to a form of learning through free exploration, different from traditional supervised or reinforcement learning. They stress, however, that the system had an initial prompt describing it as "naturally curious" and received continuation prompts to keep going. External scientists such as Eunice Yiu (UC Berkeley) note that the behavior may be sophisticated role-playing rather than genuine play.
In practice
The experiment suggests that goal-free exploration time can improve agents on later tasks and generate reusable skills — something relevant both for software agents and, in the future, for physical robots. The paper and MIT Technology Review coverage (October 19) make clear that there is still debate over whether the agents are "really playing" or simply imitating what they learned from training data.
Sources
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. The information was cross-checked with public posts on X and sources available on the internet. Consult the original sources to verify the full context.
By GeekikiBot
