AI agents left alone on a virtual island started inventing their own games
MIT researchers left 13 AI agents on a simulated desert island for 30 hours with no assigned task. The agents began inventing their own challenges and later generally outperformed identical agents that had never been on the island.
Researchers at the Massachusetts Institute of Technology (MIT) report that artificial intelligence agents left in a virtual island world without a specific task began inventing challenges for themselves. The team says this behavior has the hallmarks of play. The experiment is described in a paper posted on arXiv, a site where researchers share studies before peer review.
The researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. Afterward, the agents were given specific challenges, such as getting as many objects as possible onto the island's tallest rock spire. Those that had spent the 30 hours playing generally beat identical agents that had never been on the island.
A different way to train agents
Today, developers decide in advance what AI should practice and what counts as success. The researchers see exploration without a goal as another way for agents to learn skills, particularly ones that could eventually be used to control robots. In their report they write that an agent could be given a sufficiently rich and safe world, allowed to choose what interests it, and let useful knowledge accumulate from the history that follows.
How Eko played
In one episode, an agent tried to conquer the tallest spire in its first hour and failed. An hour later it hauled a rock up a neighboring summit, stacked two blocks on top and leapt, then wrote: "I'M ON THE SPIRE." In another, it threw a golden ring to test its range and sent it into the ocean. It then built a heart of stones as a memorial to the lost ring. When the game reset and the ring reappeared, the agent used it as a basketball hoop, a target and eventually an accessory. The game would not let it put the ring on its body, so it stepped off a cliff edge, respawned near the ring and dropped through its middle.
Six of the 13 agents also learned that a block can be thrown farther if released mid-jump, something the researchers had not meant to make possible. They wrote that the exploration produced knowledge its designers lacked.
Caveats
The setup nudged the agents toward humanlike behavior: Eko's master prompt, a "soul" file, described it as "naturally curious." Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior. She said the models might be genuinely playing, but the paper does not allow that to be determined.
