MIT researchers placed 13 AI agents powered by Claude Opus 4.7 in separate simulated environments for 30 hours with no specific objectives assigned. Left to their own devices, the agents began discovering behaviors and skills independently. One agent figured out how to stack blocks to reach an elevated structure, while others learned that releasing blocks mid-jump allowed them to throw farther than normal drops.
After the unstructured exploration period, researchers assigned specific tasks to the agents. Those who had spent time experimenting in the virtual environment performed noticeably better than agents without that prior experience, suggesting that self-directed play and experimentation improve learning and problem-solving ability. The discovery aligns with how humans and animals learn through exploration before tackling formal goals.