MIT leaves a coding agent alone on a digital island for 30 hours to study if machines play
phillip_isola · x · 2026-10-07
MIT researchers (Phillip Isola, Daniela Rus, Jacob Andreas et al.) launched an "ethology of AI agents" project: they placed Eko, an embodied agent powered by a Claude Code-style coding agent, on a simulated island with only rigid-body physics and gave it no task, reward, or goal — then watched for 30 hours, repeating the run across 13 agents.
The agents spontaneously began to play: climbing hills, stacking block towers, drawing mandalas, reinterpreting sports, and even running physics experiments with a hidden mechanic (throwing the rock hard enough spawns new cubes), later reusing learned techniques to expand their capabilities.
- Eko acts via terminal commands; the world exposes 6 actions plus an observe action returning entity positions and geometries as text
- The project site offers interactive replays pairing the world replay with the agent's full transcript
- The paper examines whether the behavior meets classical criteria for play and whether play could become a mode of machine development
- Paper, code, data, and an Inspector tool are all public
More from Embodied
- 200 builders revive vintage toys as voice agents at AssemblyAI hackathon — AssemblyAI · 2026-10-07
- Robotics is having its moment: why roboticists are Renaissance generalists — lukas_m_ziegler · 2026-10-07
- PhD student wraps up robotics planning series with EMPIRIC residual world model — tomssilver · 2026-10-07
- Wearing a BCI headband like Chinese quant traders is the new productivity hack — alexbilz · 2026-10-07
- AI firm HUMXN offers free plumbing and HVAC service in Minnesota to train home robots — brunomarsjams · 2026-10-07
- Scale and Hugging Face built a new bimanual robot dataset for training VLA models — yacinelearning · 2026-10-07