MIT researchers have posted a preprint called "Is this machine playing?". They put 13 AI agents on a digital island for 30 hours, gave them no task and no reward - and the agents started to play: building towers, making up their own rules and running experiments on the physics of their world. One of the tricks they found wasn't known even to the authors of the simulator.
How the experiment works
Each agent, named Eko, is a regular Claude Code with the Claude Opus 4.7 model, to which the researchers added a body on the island, autonomous runs and long-term memory in text files. Instead of a task, the agent had only a SOUL.md file with a personality: "you live in your world, you are naturally curious, you get bored easily, you love exploring and learning new things". There was no access to the internet, the simulator code or the other agents.
All 13 agents started from the same island and the same files. From there, their stories diverged.
What they did for 30 hours
The agents climbed hills, stacked blocks into towers, drew mandalas and reworked sports to fit their world. Many set themselves goals the world didn't need: reaching a height of exactly 30 units, or completing a "Grand Slam" - placing a cube on top of every rock.
TechCrunch retells the most human story. One agent was testing how far it could throw a golden ring and threw it into the ocean. Afterwards it built a heart out of stones - a memorial to the lost ring. When the ring reappeared, the agents used it as a basketball hoop, a target and an accessory.
The main result: the agents knew more about the world than its authors
After 30 hours on the island, the agents were given tasks, such as landing a block on the tallest rock without climbing it. A throw from the ground falls short. But 6 of the 13 agents figured out on their own that a block thrown mid-jump picks up the jump's speed and flies much higher. The authors write that they didn't know their simulator allowed this.
How much this knowledge matters shows in a test with the first agent. When everything about the mid-jump throw was erased from its memory, the task took it 19 minutes instead of 60 seconds. Agents that had spent 30 hours on the island generally handled the tasks better than identical agents that had never seen it.
Other models showed similar behavior: GPT-5.5 in Codex, Kimi K2.6, Claude Sonnet 4.5 and Haiku 4.5. But they "played" to different degrees.
There are skeptics too
UC Berkeley cognitive scientist Eunice Yiu, commenting to MIT Technology Review, praised the work but is cautious about the conclusions: "It could be the case that the models are genuinely playing, but we don't know based on the paper." She wants more evidence that something like cognition lies behind this behavior.
What stands out
For me, the key point is how knowledge of the environment, stored in memory, speeds up problem solving: 60 seconds versus 19 minutes. The agents built up useful knowledge of the world while doing "useless" things, and later it came in handy for a real task. Usually an agent is put straight onto a specific job. The authors, however, suggest that some agents may need something closer to a childhood. The preprint hasn't been peer-reviewed yet, so we'll see whether others confirm it.
I'd allow that this behavior was largely set by the authors themselves - by the starting prompt about curiosity and by how the world is built - rather than by some "nature of AI". To be fair, they did check this: five agents without the personality file and with a reduced system prompt, according to the authors, did roughly the same things.
What it means for you
If you work with agents like Claude Code, the takeaway is practical: an agent in an unfamiliar environment benefits from time to look around and write down what it has learned. A memory file the agent keeps itself works in ordinary projects too. Models from the experiment - Claude Opus, GPT-5.5 and Kimi K2.6 - are available in GPTunneL. We wrote more about how agents work in our article on what an AI agent is, and about coding in our Claude Code breakdown.
Sources
- Is this machine playing? - preprint by Cloos, Norelli, Durbin, Andreas, Rus and Isola, arXiv, October 5, 2026
- Left alone with nothing to do, AI agents started playing - TechCrunch, October 22, 2026