Cambridge – October 19, 2026 -- AI agents left alone on a virtual island for 30 hours outperformed fresh copies on a specific task, according to a new MIT paper that has not yet been peer-reviewed. The agents, copies of Claude Opus 4.7 nicknamed Eko, had no instructions beyond a prompt describing them as "naturally curious."
Researchers placed the agents on a simulated island loosely based on Roblox. Eko wrote instructions that moved its avatar and explored the limits of the world. It set its own goals, including climbing the map's highest point and testing how high it could throw a block.
Agents invent games and drawings without external rewards
One agent discovered it could create new blocks by throwing and dropping a larger rock. It then arranged blocks into drawings, moving from hearts and pyramids to its own name and the constellation Cassiopeia. Another iteration devised a version of basketball and an activity resembling "the floor is lava," jumping between small blocks without touching the ground.
In a juggling example highlighted by lead author Antonio Norelli, a postdoc at MIT's Computer Science and Artificial Intelligence Laboratory, Eko set a record of 50 consecutive catches. It then canceled the record, judging the catches too quick, and restarted under stricter rules.
Play-trained agents beat fresh copies on a peak-loading task
The agents logged activities and discoveries in text files that served as long-term memory. When tested on getting as many objects as possible onto the island's highest peak, agents with 30 hours of island experience outperformed fresh copies. Deleting lines from the memory files degraded performance.
The team says the behavior shows four hallmarks of play drawn from classic work by University of Tennessee psychologist Gordon Burghardt: self-invented rules and challenges, repetition with variation, pretending objects are other objects, and goal pursuit for its own sake. The researchers frame this as a possible third training route alongside supervised learning and reinforcement learning.
Researchers and outside experts dispute whether it is genuine play
MIT researcher Nathan Cloos conceded the behavior could be role-playing, but said the play contributed to learning. Eunice Yiu, a UC Berkeley cognitive scientist, said classifying it as play would require more evidence of underlying cognition and motivation. She noted the agents may simply be imitating humans in line with the prompt.
The setup nudged the agents toward human-like behavior. A "soul" file described Eko as curious and eager to explore. Eko also often stood idle until a continuation prompt told it to keep doing whatever it wanted. MIT professor Phillip Isola, a coauthor, said the system learned from human culture and was trained by Anthropic engineers before researchers placed it on the island.
The paper points to humanoid robots as a downstream use, where goal-free exploration could help models adapt to unexpected situations. Its conclusion: "Others may need something closer to a childhood."