MIT: Claude Opus 4.7 agents left on a virtual island invent play-like challenges

MIT researchers put 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours with no specific task. The agents began inventing challenges of their own, which the researchers say bears the hallmarks of play. The experiment is described in a paper on arXiv, a site where researchers share studies before peer review.

Afterward, the researchers gave the agents specific challenges, such as getting as many objects as possible onto the island's tallest rock spire. The Eko agents that had spent those 30 hours playing generally beat identical agents that had never set foot on the island. The source gives no numeric results for the comparison.

The researchers' pitch is a different way to build agents. Today, developers decide in advance what AI should practice and what counts as success. The researchers see exploring without a goal as another way for agents to learn skills, especially skills that could one day be used to control robots. In their report they write: "We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress. Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows."

The article walks through what one Eko did. In its first hour it wrote "Let me observe my surroundings" and set out to conquer the island's tallest spire. After failing, it came back an hour later with a backup plan: it hauled a rock up a neighboring summit, stacked two blocks on top of it, and leaped. "I'M ON THE SPIRE," it wrote.

Its next big moment went worse. Testing how far it could throw a golden ring, it sent the ring straight into the ocean. "Ouch. Lost the Ring," it wrote, before deciding to "pivot to something achievable and meaningful": building a heart made of stones, "a memorial to the lost Ring."

When the game reset and the ring reappeared, Eko used it as a basketball hoop, a target and, eventually, an accessory. The game would not let Eko put the ring around its own body, so Eko found a workaround: falling off the edge of the cliff would make it respawn in the same area as the ring. It stepped off and dropped back in, right through the middle of the ring. "I'm wearing the Ring," it announced.

The stunts were not just party tricks, the article says. Six of the 13 agents also learned they could throw a block farther by letting go of it mid-jump, something the researchers had not meant to make possible. The researchers wrote that the exploration produced "knowledge of the world that its designers lacked."

The setup nudged the agents toward humanlike behavior. Eko was never told to invent games, but its master prompt, a "soul" file, described it as "naturally curious."

Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review she was impressed but wants more evidence of the cognition behind the behavior. "It could be the case that the models are genuinely playing, but we don't know based on the paper," she said. The report itself concludes: "Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood."

Key facts

• 13 AI agents called Eko, each running on Claude Opus 4.7, spent 30 hours alone on a simulated desert island with no specific task.
• The agents invented their own challenges, which MIT researchers say bears the hallmarks of play.
• Eko agents that had spent those 30 hours on the island generally beat identical agents that had never been there on later specific challenges; no numeric results are given.
• Six of the 13 agents learned to throw a block farther by letting go of it mid-jump, something the researchers had not meant to make possible.
• The work is an arXiv paper, shared before peer review; UC Berkeley cognitive scientist Eunice Yiu wants more evidence that the models are genuinely playing.

Why it matters

Agents are normally built by deciding in advance what they should practice, what they should optimize and what counts as progress. This experiment tests the opposite: give an agent a rich and safe world, let it choose what interests it, and see whether useful knowledge accumulates. The researchers say the exploration produced knowledge of the world that its designers lacked, such as a way to throw a block farther that they had not meant to allow. They also see goal-free exploration as a route to skills that could one day be used to control robots.

Who it affects

Mainly people who design and train AI agents, and those interested in robot control, which the researchers name as a possible later use of the approach. The source describes a research experiment, not a product, and says nothing about users.

How to use it

There is nothing to install or buy. The source describes an experiment, not a tool or service. The idea on offer is a training approach: a rich, safe environment and an open-ended prompt in place of a fixed task. The paper is posted on arXiv. The researchers do not name a timeline for any practical use, beyond saying robot control could come "one day."

How solid is it

The paper is on arXiv, where studies are shared before peer review, so it has not been through that process as far as the article says. The result is stated only as agents that played generally beating identical agents that had not; no win rates or scores are given. The sample is 13 agents on one simulated island. The framing as play comes from the MIT researchers, who say the behavior bears the hallmarks of play. The article reports this as their interpretation, not a settled finding.

Risks and caveats

Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review that the models might be genuinely playing, but that the paper does not show it, and she wants more evidence of the cognition behind the behavior. The article also notes that the setup did a lot to nudge the agents toward humanlike behavior: Eko was never told to invent games, but its master prompt, a "soul" file, described it as "naturally curious." That makes it harder to say how much of the play came from the agents and how much from the prompt.

“Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.”

— The MIT researchers, in their report