MIT lets a Claude-powered agent freely "play" on a simulated island for 30 hours — and finds play expands its capabilities

LuizaJarovsky · x · 2026-10-18

An MIT paper (authors include Jacob Andreas, Phillip Isola, Daniela Rus) studies what AI agents do with no instructions at all: a Claude-powered coding agent was placed on a small simulated island and left to explore for 30 hours.

The agent, which named itself Eko, climbed hills, stacked blocks, drew mandalas, invented bowling and basketball games, and eventually used blocks to lay out constellations like Orion and the Southern Cross on the ground. The paper's findings:

Their embodied agent architecture, dubbed Eko, spontaneously generates its own activities in an unknown world;
These activities exhibit behavioral signatures of play;
Engaging in play triggers a form of learning that expands the agent's later capabilities.

The authors frame this as an alternative training regime: give an agent a rich, safe world, let it choose what is interesting, and let useful knowledge accumulate — "some capabilities are best taught by tasks; others may need something closer to a childhood." Replays of the full 30 hours are on the project website.

Commentator Luiza Jarovsky adds a governance angle: as labs automate AI R&D and after incidents like the OpenAI/Hugging Face hack, labs have overestimated how much they know about agentic behavior — making experimental studies of agent behavior essential for regulation and safety.

Original post →

MIT lets a Claude-powered agent freely "play" on a simulated island for 30 hours — and finds play expands its capabilities

More from AGI Musings