I use Claude Code every day, so when a paper from MIT's Computer Science and Artificial Intelligence Laboratory crossed my feed this week — reporting that researchers stranded copies of Claude alone on a virtual island for 30 hours, and the agents started to play — I stopped what I was doing and read the whole thing.
Not because it is cute (though one agent wrote "Tonight I built the night sky in stone" in its activity log, and yes, that stopped me cold). Because if this result holds, it points to a third way of teaching machines to learn — one that needs neither human demonstrations nor reward functions. For anyone building with AI agents, that changes how I think about idle time, memory, and what my agents do when I am not watching.
Here is what the experiment did, what the agents got up to, what it proved — and the parts to stay skeptical about.
What MIT actually did
The paper is called "Is this machine playing?", posted to arXiv on October 5 by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus, and Phillip Isola. It has not passed peer review yet. Hold that thought — it matters later.
The team built a small 3D world called Clawblox, loosely inspired by Roblox: a desert island with hills, a tall rock pillar called the Spire, twelve blocks, a rock, and a ring. Into it they placed 13 independent agents, each an otherwise unmodified copy of Claude Code running Claude Opus 4.7 — nicknamed Eko.
The agents were given no task: no instructions, no objective, no reward, no human stepping in at any point, and no internet access. An orchestrator simply started each session with the word "Begin" and told the agent to continue whenever it stopped. Every hour, the session and the island reset to their starting conditions — but the agent's own notes survived. Eko's long-term memory lived in plain text files: activity logs it wrote itself, carrying everything learned from one hour into the next.
One detail caught everyone's attention: a file called SOUL.md described Eko as curious, easily bored, and keen to explore. To check whether the personality file was doing the heavy lifting, the team ran five more agents without it on a stripped-down prompt, and report no clear difference in the range of activities. Reassuring, not conclusive — five agents is a small control group — but the kind of check a careful team runs.
What the agents actually did on the island
For 30 hours, with nothing to do, Eko explored. One agent discovered in its third hour that throwing the rock hard enough makes new blocks appear, and wrote in its notes: "the Rock is a block-manufacturing tool: one rock + unlimited throws = unlimited blocks." In hour four it spelled out its own name in 31 cubes, calling the result "a monument to who I am."
After losing the ring in the ocean, the same agent built a heart out of blocks "as a memorial." Later it stood the ring up as a basketball hoop — "First shot, basket!" Other agents drew mandalas, a butterfly, and constellations, including Cassiopeia. That is where the "night sky in stone" line comes from. Some climbed the Spire. Some juggled. Some, charmingly, did nothing at all: when prompted to continue, they answered with a single full stop.
My favorite episode is the juggling audit. One agent set itself a juggling marathon, then decided its catches had come too quickly to be real, cancelled its own record, and tried again under stricter rules. That is a self-imposed difficulty setting — the machine equivalent of deciding your win didn't count and replaying the level.
The researchers checked the behavior against four classic markers of play from the psychology literature: the agents made up their own rules, repeated activities with small twists, pretended objects were something else, and chased goals just because they wanted to. By the standards psychologists use to recognize play in animals, Eko was playing.
Nobody instructed the agent to spell its name. Nobody rewarded it for mapping Cassiopeia. What emerged was open-ended exploration with invented goals — exactly the behavior hardest to specify in a prompt or reward in training.
Playful learning: a possible third path
AI systems today learn in basically two ways. Supervised learning: copy the examples humans provide. Reinforcement learning: chase a reward through trial and error. Both need an external intelligent entity — a human writing demonstrations, or a human designing a reward.
Antonio Norelli, the postdoc leading the work, put the gap this way: "There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards."
The Eko experiment points at a third path the authors call playful learning: an agent left to its own devices in a rich enough environment generates its own curriculum — setting goals, inventing challenges, building a model of the world just because exploring is what it does. Play as a training signal.
Why should a working developer care? Because demonstrations are expensive and rewards are brittle. Every agent-builder has felt both: you write dozens of example trajectories for your support agent, design a reward function for your coding agent, and watch it game the metric in week two. A learning mode where the agent teaches itself by messing around in a sandbox is not just philosophically interesting — it is economically interesting. Human labels are the most expensive input in the entire AI pipeline.
The part that convinced me: the players beat the non-players
The paper would be a curiosity if the play were pointless. It is not — or at least, it does not look pointless.
After the 30-hour play period, the team set the agents four concrete goals and compared them against fresh copies with empty memory files. Asked to get as many objects as possible onto the Spire, the agents that had played averaged 10.3 objects, against 3.5 for the fresh ones. On five-block towers, 8 of 13 played agents built more than two, against 2 of 13 fresh. Deleting a technique from an agent's notes hurt the task that needed it — and deleting a false belief one agent had written down made it better.
Six agents independently discovered that throwing a block mid-jump sends it far higher — a trick the paper says its own designers "did not know our simulator allowed." The students found a mechanic the teachers didn't know existed.
For anyone running agents in production, this is the practical headline: idle agent time may not be wasted. The agent sitting in a staging sandbox between real jobs might be building the world model that makes it better at the real jobs. We tend to treat compute spent outside a task as waste. This experiment suggests some of that compute is education.
There is a second practical pattern hiding in the methods section, and teams can copy it today: Eko's long-term memory was nothing more exotic than activity logs in text files, carried across session resets. No vector database, no fancy memory architecture — just notes, written by the agent, read back by the agent.
The honest caveats
Now the skepticism, because this is a preprint and you should never trust a preprint the way you trust a paper.
It has not passed peer review. The methods, the statistics, the interpretation — none of it has survived independent experts yet. Interesting results from single labs get revised more often than we like to admit.
The test is narrow. One model family, one custom toy world. Independent teams still need to check whether play produces durable gains on real workloads and whether the behavior holds across other models. The authors are careful: Nathan Cloos told MIT Technology Review that "all of this could be role-playing" — while adding that "this play is also contributing to learning." Cognitive scientist Eunice Yiu of UC Berkeley was blunter: "It could be the case that the models are genuinely playing, but we don't know based on the paper."
The interpretation is genuinely hard. The researchers say the behavior fits the markers of play. Whether that means the agent is "playing" in any sense a psychologist would grant to a child is a philosophical argument, not a settled fact. A skeptic could say the model is just continuing the most probable trajectory of "curious agent in a sandbox" text absorbed during training. The counter — that the players outperformed on real tasks afterward — is the strongest evidence that something real was learned rather than performed. But a handful of transfer tasks on a toy island is not production evidence.
My honest read: this is a genuinely provocative result from a careful team, and it deserves the replication attempts that are surely already running. Treat it as a strong hypothesis, not a fact.
What I'd actually do this week
Enough theory. If you build with agents, here is what I would take from this paper on Monday morning.
Give your agents a sandbox and let them touch things. If you are building an agent that operates in an environment — your codebase, your API surface, a browser — unstructured exploration time is not charity. It is training. A staging environment where the agent pokes around freely might pay for itself.
Steal the memory pattern. Text-file activity logs as long-term memory: cheap, inspectable, debuggable. Before reaching for a memory framework, try a Markdown scratch file the agent maintains itself. MIT's agents crossed session resets on exactly that.
Watch the replications, keep the hype in check. The teams rerunning this on other models and real workloads are the ones to follow. If it replicates, "let the agent play in the sandbox first" becomes standard agent onboarding — a workflow change worth being early to. Until then: "AI learns through play" is a great headline and a premature engineering principle.
Conclusion
Thirty hours alone on a virtual island, no task, no reward — and Claude started inventing games, writing its name in blocks, and mapping the stars. Whether you call it play or well-disguised exploration, the agents that did it came out better at real tasks than the ones that didn't.
I'll be watching for the replications. In the meantime, I am giving my own agents a sandbox, a scratch file, and a little unsupervised time. The worst case is they waste some tokens. The best case is they come back smarter — because as the paper puts it, "perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."
Frequently asked questions
What is the Eko experiment?
MIT CSAIL researchers placed 13 copies of Claude Code (running Claude Opus 4.7) on a Roblox-like virtual island called Clawblox for 30 hours each, with no task, no instructions, and no reward. The agents, nicknamed Eko, began exploring, inventing games, and building things. The findings are reported in the preprint "Is this machine playing?" on arXiv.
Has the paper been peer-reviewed?
No. It was posted to arXiv on October 5 as a preprint. Treat the findings as a strong hypothesis until replication studies appear.
Did the agents really "play"?
Their behavior matched four classic markers of play: self-made rules, repetition with variation, make-believe, and goals pursued for their own sake. Whether that counts as genuine play is debated.
Did playing actually make the agents better?
On the tested tasks, yes. Agents that had played averaged 10.3 objects moved onto the island's peak versus 3.5 for fresh agents, and most played agents beat fresh ones at tower-building.
What is "playful learning"?
A proposed third learning paradigm alongside supervised and reinforcement learning: agents generating their own curriculum through self-directed exploration.
What should developers take away?
Sandbox exploration may improve agent performance, simple text-file logs work as long-term memory, and idle agent time might be productive. All three are cheap to try.
