Claude agents invent games and improve at virtual island tasks

Researchers placed copies of Claude Opus 4.7, nicknamed Eko, on a simulated desert island for 30 hours without explicitly instructing them to invent games or explore. The agents developed their own challenges, arranged blocks into drawings and constellations, and devised activities resembling basketball and “the floor is lava.” Their behavior suggests that playful exploration could offer a learning approach alongside supervised learning and reinforcement learning.

The agents recorded discoveries in text files that served as long-term memory. When later assigned tasks, including moving objects onto the island’s highest peak, experienced agents outperformed fresh copies. Deleting parts of their memories reduced performance. In another example, an agent recorded 50 consecutive catches while juggling, then rejected its achievement and imposed stricter rules because the catches had happened too quickly.

The findings have not been peer-reviewed, and researchers caution against equating the behavior with human or animal play. A guiding prompt described Eko as naturally curious and eager to explore, while continuation prompts restarted agents that stopped acting. UC Berkeley cognitive scientist Eunice Yiu said behavioral evidence alone cannot establish the cognition or motivation behind play. The experiments nevertheless suggest that open-ended exploration can contribute to useful skills, potentially helping future agents adapt to unfamiliar environments.