Researchers from MIT introduced Eko, an embodied coding agent capable of writing code and interacting within a virtual environment. During 30-hour sessions, the agent operated on a virtual island without specific tasks, rewards, or predefined goals. It engaged in activities such as climbing terrain, building structures, drawing, and testing physical interactions.
The study involved 13 agents placed in a sandbox containing blocks, stones, and rings. The agents demonstrated self-directed behaviors, including stacking blocks, climbing, creating patterns, and inventing sports rules. Each agent developed a unique behavioral history, although some patterns were repeated across the group.
The authors emphasize that the research does not suggest the agents possess consciousness or experience pleasure. The observations are limited to actions and stored knowledge within the simulation. The preprint has not yet undergone peer review, and the researchers note that the environment was intentionally restricted.
The team argues that narrow benchmark tests often overlook emergent behaviors that appear when AI agents have time to explore and refine strategies. The study raises questions about the reproducibility of these results, noting that detailed documentation of prompts, memory, and safety boundaries is required before assessing potential real-world applications.
