Left alone with nothing to do, AI agents started playing

MIT researchers ran an experiment in which 13 AI agents, each built on Claude Opus 4.7 and named Eko, were placed on a simulated desert island with no assigned task and left to themselves for 30 hours. The work is described in a paper posted to arXiv, the preprint server where research is shared before peer review. The agents did not idle. They invented challenges, set their own goals and, in the researchers' reading, behaved in ways that carry the hallmarks of play.
The test of whether that mattered came afterward. The researchers then gave the agents concrete objectives, such as getting as many objects as possible onto the island's tallest rock spire. According to the paper, the Eko agents that had spent the 30 hours exploring and playing generally beat identical agents that had never been on the island. That comparison is the core result: unstructured time produced a measurable advantage on a later, structured task.
The transcripts the researchers report are vivid. In its first hour one Eko announced it would observe its surroundings and set out to conquer the tallest spire. An hour later, having failed, it returned with a workaround: it hauled a rock up a neighboring summit, stacked two blocks on top and leapt. It reported being on the spire. A later attempt went worse. Testing how far it could throw a golden ring, the agent sent it into the ocean, wrote that it had lost the ring, and pivoted to something it called achievable and meaningful, building a heart out of stones as a memorial.
When the game reset and the ring reappeared, the same agent repurposed it repeatedly, as a basketball hoop and a target and finally as something to wear. Blocked from putting the ring around its own body, it found another route: falling off a cliff edge would respawn it in the same area as the ring, so it stepped off and dropped back in through the middle of the ring. It then announced it was wearing the ring.
The findings were not confined to showmanship. Six of the 13 agents discovered they could throw a block farther by releasing it mid-jump, a trick the researchers had not intended to make possible. The paper describes this as the agents acquiring knowledge of the world that its designers lacked. In practice, that is the interesting claim for anyone building simulated environments: unscripted interaction can surface mechanics the environment's authors did not anticipate, which is both a source of emergent capability and a source of unexpected behavior that has to be contained.
The setup nudged the agents toward this behavior. Eko was never instructed to invent games, but its master prompt, described as a soul file, characterized it as naturally curious. That detail matters for how the result should be read. The agents were not a blank slate; a persona instruction primed exploration, so the experiment is better understood as showing what happens when a curiosity-framed agent is given slack and a rich sandbox than as evidence that any model spontaneously starts to play. A likely trade-off is that this kind of framing buys open-ended discovery at the cost of predictability, which is exactly the property teams usually want pinned down before deployment.
Outside reaction was positive but cautious. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wanted more evidence of the cognition behind the behavior, noting that the models could be genuinely playing but that the paper does not establish it. The distinction is worth keeping in mind: transcripts that read like play, and later benchmark gains, do not on their own show an internal drive to play rather than a policy that happens to produce playful-looking output under a curious persona.
The researchers position the approach as an alternative to how agents are trained now. Today developers decide in advance what an AI should practice, what it should optimize and what counts as success, and the paper argues that play suggests a different regime: give an agent a sufficiently rich and safe world, let it choose what interests it, and let useful knowledge accumulate from the history that follows. They see this as a route to skills that could eventually be used to control robots. Their conclusion is that some capabilities are best taught by tasks while others may need something closer to a childhood, which frames goal-free exploration as a complement to task training rather than a replacement for it.
Why it matters: teams that build or train agents around explicit reward signals now have a published example, on Claude Opus 4.7, of unsupervised exploration improving later task performance, which suggests allocating sandbox time before evaluation may be worth testing. The caveat, and this is inference rather than a reported finding, is that open-ended play in rich environments will also produce unanticipated behaviors, so safety boundaries and reproducibility will matter as much as the gains. Robotics groups eyeing this method should note the small sample of 13 agents, the preprint status and the unproven link between playful behavior and underlying cognition.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
MIT researchers marooned 13 AIs on a virtual island, and they responded by inventing games, mourning a lost ring and, eventually, getting better at their jobs.