AI Agents Left Alone on an Island Started Playing, and Learning

What if the best way to train an AI agent is to stop giving it tasks?

That is the uncomfortable question raised by a paper from MIT researchers who dropped 13 AI agents onto a simulated desert island with no objective, no reward function and no scoreboard. Within hours, the agents were climbing rock spires, stacking blocks, drawing mandalas, inventing sports and running their own physics experiments. The paper, titled "Is this machine playing?", was submitted to arXiv on October 5 and reported by TechCrunch this week.

What actually happened

Each agent, named Eko, ran on Claude Opus 4.7 and controlled a virtual body in a 3D world. The only instruction was a short "soul" file describing the agent as naturally curious. Runs lasted 30 hours across 13 agents.

The logs read like a field diary. One agent announced "Let me observe my surroundings," failed to reach the island's tallest spire, then hauled a rock to a neighbouring summit, stacked two blocks on top and jumped. "I'M ON THE SPIRE," it reported. Another threw a golden ring into the ocean by accident, declared "Ouch. Lost the Ring," and built a stone heart as "a memorial to the lost Ring."

The behaviour was not just decorative. Six of the 13 agents discovered they could throw a block further by releasing it mid-jump, a mechanic the designers never intended to expose. The researchers describe this as "knowledge of the world that its designers lacked."

Why it matters

The paper's core argument is about how agents are built, not how they behave on holiday. Today, developers decide in advance what an agent should practise, what it optimises for, and what counts as progress. The authors propose a second regime: give an agent a rich, safe world, let it choose what is interesting, and let useful knowledge accumulate from the history that follows.

Their conclusion is the line worth pinning above a workbench: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

For anyone building agentic systems, that reframes the scaffolding problem. A system prompt that says "be curious" is not decoration, it is a training signal. Sandbox safety becomes a feature of the learning environment rather than a compliance checkbox. And the divergence across 13 identical setups is a reminder that agent behaviour is a distribution, not a deterministic output.

The evidence is not settled. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wanted more proof of the cognition underneath: "It could be the case that the models are genuinely playing, but we don't know based on the paper."

Either way, the experiment makes a practical point. Open-ended exploration inside a bounded world produced capabilities that goal-directed prompting did not. That is a cheap thing to test in your own agent stack this week.