MIT Study on Free-Exploring AI Agents Finds Play Behavior Enhances Task Skills

Researchers at the Massachusetts Institute of Technology (MIT) conducted an experimental behavioral study on an AI agent operating without explicit task directives. The team deployed an AI agent named Eko, controlled by Claude Code, into a simulated island environment with varying terrain and building blocks. With the sole prompt to maintain curiosity and continue exploring, the agent was given 30 hours of unstructured time per round across 13 iterations. The findings revealed that Eko exhibited purposeful and spontaneous activities rather than idle wandering. In the second iteration, the agent arranged blocks into a diamond monument aligned with cardinal directions and executed perimeter walking rituals. In subsequent runs, it constructed stonehenge-like arrangements and towers, reorganizing debris into new configurations after structural collapses. The agent also drew in the sand, climbed elevated platforms, and constructed temple-like environments. The MIT team noted that this unstructured exploration led to demonstrable functional improvements, such as faster traversal to elevated areas and improved spatial navigation. Researchers suggest that curiosity-driven play allows autonomous agents to acquire and refine skills applicable to subsequent tasks. The underlying code for the experiment has been open-sourced.

Key takeaways

• MIT researchers tested an AI agent across 13 thirty-hour rounds of open exploration on a virtual island, observing autonomous building, naming, and traversal behaviors.
• The experiment confirmed that the agent improved its climbing speed and navigation over time, demonstrating skill acquisition through unstructured exploration.
• The research team framed curiosity-driven play as a viable pathway for AI skill acquisition and open-sourced the project codebase.