AI Agent Stranded on Virtual Island — and It Started Playing

Researchers from MIT CSAIL (Nathan Klus, Antonio Nori, Daniel Durbin, Jacob Andreas, Daniela Rus, Phillip Isola) in the paper “Is this machine playing?” (arXiv:2610.07130) placed an LLM coder agent as the “mind” of a virtual body on a digital island — without tasks, rewards, or prescribed activities. Over 30-hour sessions, 13 agents (nicknamed Eko) spontaneously transitioned to play: building towers, mandalas, and a Stonehenge-style monument, while also conducting their own physical experiments.

Compiled by Sergey KostenchukPublished 2026-10-17Updated 2026-10-17

2026-10-17 Research

🤖 AI Agent Stranded on Virtual Island — and It Started Playing

MIT CSAIL (“Is this machine playing?”, arXiv:2610.07130) placed an LLM agent as the “mind” of a virtual body on a digital island without tasks, rewards, or prescribed activities. Over 30 hours, 13 Eko agents built towers, mandalas, and a Stonehenge-style monument, and conducted their own physical experiments: a block thrown during a jump travels farther.

🌍 For the industry, this is an empirical observation: an agent in a safe environment without goals spontaneously plays and accumulates knowledge that the developer did not program. Unsupervised research in a “sandbox” could complement task-based learning and RL, but this is currently just one experiment.

👤 The paper (13 pages) is open on arXiv: you can see what the AI is doing when left “alone” for 30 hours. It’s a reason to think about what “play” means for a machine.

Source 1: https://arxiv.org/abs/2610.07130 Source 2: https://gizmodo.com/researchers-stranded-an-ai-agent-on-a-virtual-island-things-got-very-weird-2000825998

Sources