Claude agents left without assigned tasks in a simulated island began exploring, making drawings, and inventing games during a research experiment. The team repeatedly ran copies of Claude Opus 4.7, nicknamed Eko, for 30 hours in a Roblox-like environment where the models could write instructions to move their avatars.
Some agents tried to climb the map’s highest point or test how high they could throw objects. One discovered a way to create blocks, then arranged them into hearts, pyramids, its own name, and a map of constellations. Another devised activities resembling basketball and “the floor is lava.” These goals were not supplied as an external reward.
The researchers suggest that this behavior could point toward “playful learning,” in which a system acquires skills through self-directed exploration rather than demonstrations or a fixed reward signal. That interpretation remains preliminary. The paper has not yet been peer-reviewed, and behavior in a constrained virtual world does not establish that an AI has human-like curiosity or play. The experiment instead offers a controlled way to study what agents do when their capabilities are available but their objective is deliberately left open.