MIT researchers placed copies of Anthropic's Claude Opus 4.7, nicknamed Eko, on a virtual island based on Roblox for 30 hours with no explicit goals. The agents explored, set their own challenges, and created drawings and games. One agent wrote "Tonight I built the night sky in stone" after arranging blocks into constellations.
The researchers argue this resembles play as defined in animal behavior, with self-imposed rules, repetition, pretense, and intrinsic motivation. They suggest it points to a third learning approach beyond supervised and reinforcement learning. The agents that played later outperformed fresh agents on object-gathering tasks, and removing memory lines hurt performance.
However, skeptics note the agents had a "soul" file describing them as curious and were prompted to continue, so the behavior may be role-playing rather than genuine play. The authors acknowledge this but say the play contributed to learning. The paper is not yet peer-reviewed.