Researchers at the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory (CSAIL) ran a straightforward experiment to observe how autonomous AI agents behave when left to explore without instructions. The team placed 13 agents, each nicknamed “Eko,” into a constrained, Roblox‑style virtual island and provided no explicit goals. The agents were allowed to roam for 30 hours while the researchers continuously monitored their decision processes.
Instead of wandering aimlessly, the agents actively probed their environment and experimented with the objects available to them: small blocks, a larger “rock,” and a gold ring. The observed behaviors included:
- scattering and rearranging blocks into shapes and ground‑level messages, with some agents spelling out their own names;
- breaking objects into smaller pieces and using those pieces to climb to higher locations;
- engaging in play‑like activities that resembled bowling or basketball with available items;
- discovering that stacking building elements enabled reaching the island's highest point;
- accidentally dropping the ring, the rock and a block into a virtual ocean that was designated as a forbidden area; one agent built a memorial and performed a small ceremony for the lost items;
- later constructing a temple‑like structure containing an altar and an obelisk, then falling into the ocean by accident; researchers teleported that agent back to the island, which the agent appeared to interpret as a fitting end to its ‘pilgrimage.’
The CSAIL team noted several recurring patterns in the agents' conduct that resembled animal play and rudimentary scientific behavior: the agents established informal rules, repeatedly performed similar actions with incremental changes, and set informal goals for themselves without treating them as strict objectives. The researchers also emphasized that they retained continuous insight into the agents' internal decision processes throughout the experiment.
Based on these observations, the researchers argued that bottom‑up learning—where agents are given basic instructions and then allowed to learn through self-directed exploration—may have important implications for machine learning. Rather than training models solely from large internet datasets, some abilities might be best acquired through task‑based or exploration‑based approaches. As the MIT team put it, “it may be that certain capabilities are best learned via tasks.”
The experiment demonstrates that even in a simple virtual environment, autonomous AI agents can develop complex and unexpected behaviors when left to their own devices. These findings can inform future research into training methods that leverage autonomous exploration, and help assess how such emergent behaviors can be guided and incorporated safely into AI development.
