What do AI agents do when nobody gives them a task? According to a report by MIT Technology Review, the result looks surprisingly like play. Researchers repeatedly placed copies of Claude Opus 4.7, called Eko, alone on a virtual island for 30 hours.
What the agents did
The island was loosely modeled on the video game Roblox. The agents moved their avatars using instructions they wrote themselves. They explored the edges of the world, tried to climb to the highest point, or tested how high they could throw a block.
One agent arranged blocks on the ground to make pictures: first hearts and pyramids, later its own name and even constellations. Another invented games, including a kind of basketball and a variant in which you have to jump between small blocks without touching the ground. While juggling, one agent set a record of 50 catches in a row, then declared it invalid because the catches had been too fast, and started over with stricter rules.
Playful learning as another approach
AI systems mostly learn by imitating human examples or through reinforcement learning, that is, trial and error aimed at a reward. The experiment points to another path. The agents recorded their experiences in text files that served as long-term memory.
In later tasks, such as bringing as many objects as possible to the highest peak, agents with play experience performed better than fresh copies. When lines were deleted from the memory files, performance dropped. However, the study has not yet been reviewed by independent experts.
Play or just imitation?
The authors concede that it could all be role-play. But they stress that playing demonstrably contributed to learning. Cognitive scientist Eunice Yiu finds the result impressive, but misses evidence of thinking and motivation behind it.
The setup also steered the agents: a text file described Eko as curious and eager to explore. In addition, the agent often stood around idle until a prompt to continue nudged it along. Whether it was genuine play remains open. In practical terms, free exploration could help when language models in robots encounter the unexpected.
