MIT researchers say AI agents left on a virtual island with no task began inventing their own games. The team says the behavior shows signs of play.
The study was posted on arXiv, a site where researchers share work before peer review. TechCrunch reported on the paper on October 22, 2026.
How the Island Experiment Worked
The researchers placed 13 AI agents called Eko on a simulated desert island for 30 hours. Each agent ran on Claude Opus 4.7, a model made by Anthropic.
The agents were not given a goal. They were free to explore the island and decide what to do.
After the 30 hours, the researchers gave the agents set challenges. One task asked them to get as many objects as possible onto the tallest rock spire.
The Eko agents that had spent time playing generally beat identical agents that had never been on the island. Today, developers usually decide in advance what an AI should practice and what counts as success.
The researchers say playing without a goal could be another way for AI agents to learn skills. That includes skills that might one day help control robots.
“We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress,” the researchers wrote. They said play lets an agent pick what is interesting in a safe world.
What the Agents Did on Their Own
“Let me observe my surroundings,” one Eko agent wrote in its first hour. It then tried to climb the island’s tallest spire and failed.
An hour later, it came back with a new plan. It hauled a rock up a nearby summit, stacked two blocks on top, and leapt. “I’M ON THE SPIRE,” it wrote.
The same agent later tested how far it could throw a golden ring. It threw the ring into the ocean and wrote, “Ouch. Lost the Ring.”
It then built a heart out of stones as “a memorial to the lost Ring.” When the game reset and the ring returned, the agent used it as a hoop, a target, and an accessory.
The game would not let Eko place the ring around its body. The agent found that falling off a cliff made it respawn near the ring, so it stepped off and dropped through it. “I’m wearing the Ring,” it said.
Six of the 13 agents also learned they could throw a block farther by letting go mid-jump. The researchers had not meant to make this possible. They wrote that the exploration produced “knowledge of the world that its designers lacked.”
The setup did push the agents toward humanlike behavior. Eko was never told to invent games, but its master prompt, called a “soul” file, described it as “naturally curious.”
Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Technology Review on October 19 that she was impressed but wants more evidence. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” she said.
The researchers framed the work as a new way to build agents. “Perhaps some capabilities are best taught by tasks,” they wrote. “Others may need something closer to a childhood.”
