Left to explore the virtual island without a specific task, the AI agent began inventing its own challenges. The MIT researchers say this behavior has characteristics of play.
The researchers left 13 AI agents called Eko, each running Claude Opus 4.7, alone on a simulated desert island for 30 hours. They describe their experiment in a paper posted on arXiv, a site where researchers share their work before peer review. They then gave the agents specific challenges, such as getting as many objects as possible onto the island’s tallest rock spire. Eko agents who spent 30 hours playing generally defeated identical agents who had never set foot on the island.
Now, developers decide in advance what AI should practice and what is considered success. The researchers see this technique of goalless exploration as another way for AI agents to learn skills, especially skills that can one day be used to control robots.
“We typically develop artificial agents by deciding in advance what to practice, what to optimize, and what to consider as progress,” the researchers wrote in their report. “Play suggests a different regime: give agents a sufficiently rich and secure world, let them choose what is interesting, and accumulate useful knowledge from subsequent history.”
“Please let me observe the surroundings,” one Echo wrote during the first hour of arriving on the island. Ambitiously, it set out to conquer the highest spire on the island. But an hour after that failure, he returned with an alternative plan: carry the rock to the next peak, stack two blocks on top of it, and jump off.
It said, “I’m in the Spire.”
The next big moment for the agent was even worse. When he tried to see how far he could throw his gold ring, it was rolled up and thrown straight into the sea. “Oh, I lost my ring.” The magazine wrote, before deciding to “pivot to something achievable and meaningful,” building a heart made of stone, “a memorial to the lost ring.”
When the game reset and the ring reappeared, Eco quickly made it function as a basketball hoop, target, and eventually an accessory. When Echo was unable to put a ring around her body in the game, she came up with a workaround. The idea was that if you fell off the edge of a cliff, you would respawn in the same area as the ring. It stepped off the edge and fell back into the middle of the ring.
“I’m wearing a ring,” he announced.
Stunts like this weren’t just party tricks. Six of the 13 agents also learned that they could throw the block farther if they let it go mid-jump. This was something the researchers never intended to make possible. The expedition provided “knowledge of the world that the designers lacked,” the researchers wrote.
This setting was very helpful in guiding the agent towards human-like behavior. Mr. Eco was never asked to invent a game, but his master prompt, the “Soul” file, described it as “naturally curious.”
Eunice Huy, a cognitive scientist at the University of California, Berkeley, told MIT Tech Review that she was impressed, but wanted more evidence of the cognition behind behavior. “It’s possible that the models are really playing, but we don’t know based on the paper,” she says.
Researchers frame this as another way to build agents. “Perhaps some functions are best learned by doing tasks,” their report concludes. “Some people may need something closer to their childhood.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
