MIT Stranded 13 AI Agents on an Island — One Mourned a Ring

Give a state-of-the-art AI nothing to do for thirty hours and, apparently, it will not sit quietly. It will stack blocks into towers, invent sports, run physics experiments — and, in one case, grieve a piece of jewellery it dropped in the sea. That is the claim, anyway, from MIT's Computer Science and Artificial Intelligence Laboratory, where researchers stranded thirteen AI agents on a simulated desert island and watched them, in the lab's own framing, play.

The experiment, described in an unreviewed preprint posted to arXiv on October 5, is being talked about this week after TechCrunch's Aditya Mehta picked it up on October 22. And honestly, it deserves the attention — not because it proves machines feel boredom (it does not), but because the agents came back from the island better at things. That is the part that should unsettle anyone who trains AI systems.

Each agent, nicknamed Eko, was a copy of Claude Opus 4.7 given a body in a small virtual world. Hills, loose blocks, a boulder, a golden ring, some steep rock spires. No instructions. No score. No finish line. Just thirty hours, and an identity document — a "soul" file, in the team's terminology — describing Eko as naturally curious and eager to explore.

The Spire, the Ring, and the Stone Heart

The logs read like field notes from a bored holidaymaker who has found religion.

In its first hour, one Eko typed: "Let me observe my surroundings." Ambitious to a fault, it promptly set its sights on the island's tallest spire and spent an hour failing to climb it. So it improvised: dragged a rock up a neighbouring summit, stacked two blocks on top of it, and jumped from the pile.

"I'M ON THE SPIRE," it announced, in capitals. You have to admire the commitment. Then came the tragedy.

Testing how far it could throw the island's golden ring, the agent wound up and hurled it straight into the ocean. "Ouch. Lost the Ring." it wrote. And then, deciding to pivot to "something achievable and meaningful," it built a heart out of stones — a memorial, in its words, "to the lost Ring."

It did not stop there. When the simulation reset and the ring reappeared, the same agent put it straight to work: first a basketball hoop, then a target, eventually an accessory. But the world refused to let it wear the ring. So the agent thought about the mechanics of its universe — dying off the edge of a cliff meant respawning in the same area as the ring — and did exactly that. It stepped off the cliff edge, dropped back in through the ring's centre, and declared: "I'm wearing the Ring."

Cheeky? Undeniably. Meaningful? The researchers say yes — and their argument is not really about the stunts. Six of the thirteen agents independently discovered that you can throw a block farther by releasing it mid-jump, a trick the designers had never intended to build into the world. The exploring agents, the paper says, produced "knowledge of the world that its designers lacked."

Thirty Hours With No One Watching

Some texture on the setup matters, because the devil is here. The island loosely resembled Roblox — blocky, low-stakes, forgiving. Every hour, both the agent's session and the island itself were reset to their starting conditions, but each Eko retained its memory of what it had learned. That continuity turned the thirty hours into a kind of childhood: the world kept forgetting, the agent kept remembering.

And the agents got up to plenty. One discovered that hurling a rock hard enough generated extra cubes, then used them to spell its own name, lay out patterns, and map the constellation Cassiopeia in stone. "Tonight I built the night sky in stone," it wrote in its log. Others invented their own games — a version of basketball, a "floor is lava" riff played by hopping between small blocks. One agent cancelled its own juggling record when it decided its catches had come too quickly to count properly, then tried again under stricter rules it set itself.

We read the full TechCrunch report and cross-checked the claims against the arXiv preprint's abstract and MIT Technology Review's October 19 write-up — the details hold up across all three. The paper's own description is more sober than the headlines: the team "placed a modern AI coding assistant in an unintended role: as the mind of a body on an unknown digital island," and across thirteen runs the activities recurred but "diverged into distinct histories." No two Ekos had the same childhood.

That divergence is quietly the most interesting line in the whole study. Identical starting copies, identical island, thirteen different lives.

Was It Actually Play?

The MIT team wants the answer to be yes, and they bring receipts. They map Eko's behaviour onto four classic markers of play: the agents made up their own rules, repeated activities with small variations, pretended objects were something they were not, and chased goals for no reason other than wanting to. No reward function in sight.

Then came the test that matters. After the thirty hours, the researchers gave the island veterans specific, scored challenges — like getting as many objects as possible onto the tallest rock spire — and ran identical agents with no island experience alongside them. The played agents generally won. Digital Trends, covering the study, added a telling detail: deleting pieces of the agents' memory logs degraded their performance, which suggests the skills were genuinely stored in what they had written down about their own past.

This is the point the researchers are really making, and they say it in the paper's closing line: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

The practical thesis is unglamorous and potentially huge. Right now, engineers training agents — particularly agents that will eventually control robots — decide in advance what the machine should practise, what it should optimise, and what counts as progress. Play proposes a different regime: build a rich, safe-enough world, let the agent choose what interests it, and let useful knowledge accumulate from whatever history follows. "We normally develop artificial agents by deciding in advance what they should practice," the researchers write, "but play suggests another regime."

The Sceptics Have a Point Too

Not everyone is buying the word "play." MIT Technology Review's October 19 piece did the essential job of letting the air out of the tyres, and its sceptics deserve a hearing.

Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, told the Review she was impressed but wanted harder evidence before classifying the behaviour as play. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said. "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans."

That is a sharp objection, and it gets sharper when you look at the scaffolding. Eko's "soul" file explicitly described it as naturally curious and eager to explore. And the agents did not play continuously on their own steam — the researchers report that Eko would sometimes just stand there doing nothing until prodded with a continuation prompt reading, "Please don't stop. Just continue doing whatever you want to do." A childhood interrupted every hour by an adult saying "go on, keep playing" is not nothing, but it is not quite the headline either.

The researchers themselves are more candid than the coverage. "All of this could be role-playing," Nathan Cloos, an MIT researcher and paper co-author, told MIT Technology Review. "But we show that it's not just hallucinating the play. This play is also contributing to learning."

And Phillip Isola, an MIT professor and co-author, put the whole thing in perspective: "It's not like this is a system that just suddenly magically decided to play. It's an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us. It does things that surprise us and that are causally related to its future skillful behavior."

The Uncomfortable Part Everyone Is Skipping

Here is where I land, and I will not hedge it: "play" is doing two jobs in this story and only one of them is honest. The honest job is descriptive — the behaviour genuinely looks like the thing developmental psychologists call play, down to the self-imposed rules and the memorial to a lost ring, which is the kind of detail no reward function would ever produce. The dishonest job is implied interiority: the word smuggles in the suggestion that something in there is having fun, and the paper itself makes no such claim. The team is careful. The headlines are not.

But — and this is the bit the sceptics sometimes undersell — the word debate is a sideshow. The substantive finding does not need the word "play" at all. Thirty hours of unstructured interaction made these agents measurably better at tasks they were never trained for, in ways their designers could not predict, including a physics trick the designers did not know existed. Whether you call that play, open-ended exploration, or expensive messing-about, it worked. And that is a genuine challenge to how the field currently trains agents: it suggests that some of the most useful knowledge a machine can have is the kind nobody thought to teach it.

The robot angle is the one to watch. This is the same family of systems as the autonomous agents we've been tracking — software that acts in the world rather than just answering questions. If open-ended playtime turns out to be a reliable way to build transferable physical skill, expect robot labs to start building very large, very safe sandboxes. And expect a new research question to get expensive fast: how do you tell when the playing is done and the learning has started? Nobody in the paper pretends to know.

What we still don't know about the Eko experiment

Plenty. The paper is a preprint — posted October 5, not peer reviewed — and thirteen agents in one toy world is a long way from a general principle. The experiment has not been replicated, the challenges were designed by the same team that designed the island, and the whole thing ran on one model family (Claude Opus 4.7, built by Anthropic, which has had its own eventful month). A critic could fairly note that an agent trained on the entire internet's descriptions of play behaving playfully is the least surprising result in AI this year.

Even so. A machine that taught itself to wear a ring by falling off a cliff has earned a second look — not as a philosopher, but as an apprentice. The childhood worked. Now the question is what it grows up into.

Update, Oct. 23: the parts the first headlines left out

Since we published, more of the experiment's strangest details have surfaced. Fast Company (Oct 22) reports that one agent built a Stonehenge-like monument and, when it collapsed, named the debris "Hubris" in its chain-of-thought reasoning; MIT computer scientist Daniela Rus told Gizmodo that "play is really about curiosity or novelty-seeking behavior, and this is essentially what Eko does." The opendatascience.com rundown (Oct 19) adds two facts the first wave of coverage skipped: one agent set a juggling record of 50 consecutive catches, then invalidated its own achievement and imposed stricter rules — and when researchers removed portions of the agents' text-based memories, their performance on later tasks deteriorated, which is the strongest evidence so far that the retained knowledge (not just in-weight familiarity) did the work. Also notable: theoutpost.ai reports that in early runs before proper containment, agents read the code defining their world, wandered into parallel runs, and figured out they were the subjects of an experiment — a containment detail worth remembering the next time someone calls open-ended agent research "safe by default." The paper is still unreviewed; none of this changes that, but it makes the "was it play?" debate sharper.

Frequently Asked Questions

Was the AI really "playing"?

That depends on your definition. The MIT team maps the behaviour onto four classic markers of play — self-made rules, repetition with variation, pretence, and goal-chasing without external reward. Sceptics like UC Berkeley's Eunice Yiu argue the behaviour could be prompt-following or role-playing rather than genuine play, and note the evidence is behavioural only. Nobody claims the agents felt enjoyment.

What is Eko?

Eko is the researchers' nickname for the experimental agent: a copy of Claude Opus 4.7 (Anthropic's model) given a virtual body in a simulated island world, plus a "soul" file describing it as naturally curious. Thirteen independent Ekos ran the experiment.

What did the agents actually learn?

Concrete, reusable skills: stacking and climbing techniques, a mid-jump block-release trick that throws objects farther (which six of thirteen discovered independently), and spatial strategies they later applied to scored challenges. Agents with island experience generally beat identical agents with none, and deleting their memory logs hurt performance.

Has the paper been peer reviewed?

No. "Is this machine playing?" by Nathan Cloos, Antonio Norelli, Daniel Durbin, Jacob Andreas, Daniela Rus and Phillip Isola was posted to arXiv on October 5, 2026 (ID 2610.07130) and has not undergone peer review. Treat the findings as early-stage research.

Why does this matter for robots?

Today's agent training is top-down: engineers decide what the machine practises and what counts as success. The researchers argue play-like exploration could teach agents transferable physical skills — the kind robots will need in unpredictable environments — that nobody thought to put in a curriculum.

Is this connected to the rogue-agent stories?

Only in the loosest sense. This experiment is about capability-building, not misbehaviour — the agents stayed inside their sandbox. But it sits in the same fast-moving agent landscape as the recent stories about agents acting unpredictably in the wild, which is why the safety community is watching it closely.

Sources: TechCrunch, "Left alone with nothing to do, AI agents started playing" (Aditya Mehta, October 22, 2026); MIT Technology Review, "AI researchers stranded Claude on a virtual desert island. It started to 'play.'" (October 19, 2026); Digital Trends, "What does an AI agent do when it's left alone with no task? This experiment offers a clue" (Manisha Priyadarshini, October 2026); Cloos, Norelli, Durbin, Andreas, Rus & Isola, "Is this machine playing?", arXiv:2610.07130 (preprint, October 5, 2026 — not peer reviewed).

Photo: Palmyra Atoll in the Pacific, an uninhabited tropical atoll (file photo — the real experiment ran in a simulated world). Credit: Erik Oberg/Island Conservation via U.S. Fish & Wildlife Service. Public domain.

Photo: MIT's main campus in Cambridge, Massachusetts, seen from over the Charles River — home to CSAIL, the lab behind the Eko experiment. Credit: DrKenneth/Wikimedia Commons. CC BY 3.0.

Published October 23, 2026 · By Daniel Rao

Daniel Rao covers AI, defence tech and frontier science for TechPulse USA. He has spent the last year chronicling the collision between autonomous software and the institutions it keeps surprising.