Researchers conducted an experiment with AI agents based on Claude Opus 4.7, nicknamed Eko, by stranding them alone on a virtual desert island for 30 hours. The island was modeled loosely on the video game Roblox. During this time, the agents explored the environment and engaged in spontaneous activities without direct instructions, revealing behaviors akin to play, according to technologyreview.com.
The AI agents wrote instructions to move their avatars around the island, setting their own goals such as climbing the highest point or throwing blocks to test limits. One agent discovered it could create new blocks by manipulating a larger rock and began arranging these blocks into drawings, including hearts, pyramids, its own name, and star constellations like Cassiopeia. Another agent invented games resembling basketball and "the floor is lava," jumping between blocks without touching the ground.
This experiment suggests a novel approach to AI training beyond supervised and reinforcement learning. Instead of copying human examples or pursuing explicit rewards, the agents demonstrated a form of playfulness that led to creative exploration and goal-setting. This behavior may open new avenues for developing AI systems that learn through self-motivated interaction with their environments, expanding the understanding of AI autonomy and creativity.
The findings were published on October 19, 2026, by technologyreview.com, highlighting the potential for AI to engage in unsupervised, playful learning. The experiment lasted 30 hours, during which the Claude-based agents exhibited these emergent behaviors on the virtual island.