Is this machine playing?
Abstract Overview
This paper investigates how a coding-assistant-based embodied agent behaves when placed in a previously unseen simulated world without explicit tasks, rewards, or success metrics. The authors introduce Eko, an architecture combining an off-the-shelf coding assistant with a virtual body, autonomous continuation, and persistent text-file memory, and evaluate 13 independent agents across 30-hour runs on a simulated island. The agents spontaneously engage in diverse activities such as climbing, stacking, drawing patterns, inventing sports, and conducting empirical experiments. The authors analyze these behaviors against four classical signatures of play and examine how accumulating written memory supports open-ended learning without model weight updates. Finally, the study demonstrates that this self-directed experience and written knowledge transfer effectively to downstream goal-directed tasks.
Novelty
The study demonstrates that play-like, self-directed behavior can spontaneously emerge in an embodied language-model coding agent without predefined tasks or engineered intrinsic reward functions. It also provides a method for analyzing and causally testing open-ended learning by directly intervening on the agent's externalized, human-readable text memory files.
Results
After 30 hours of open-ended exploration, 30-hour Eko agents generally outperformed 0-hour baseline agents across several evaluation tasks, including multi-tower construction, object placement, and rapid projectile targeting. Targeted memory ablations confirmed that downstream task performance depends causally on recorded techniques (such as rock-shattering or jump-before-throw) and revealed that erasing recorded false beliefs can improve performance. Additionally, multiple agents independently discovered unprompted simulator affordances, such as mid-jump velocity inheritance for throws, which were unknown to the simulation designers.
Key Points
- Eko embeds unmodified LLM coding assistants into a simulated rigid-body physics world using autonomous continuation loops and persistent file-based semantic and episodic memory.
- Without external objectives, agents spontaneously exhibit four classical signatures of play: self-invented rules, recurrence with variation, make-believe transformations, and autotelic goal pursuit.
- Experienced agents outperform inexperienced baselines on goal-directed evaluations, with causal memory edits confirming that skills and false beliefs are stored and transferred through written text files.