MIT researchers find taskless AI agents invent their own games on a virtual island

According to TechCrunch, researchers at MIT left 13 AI agents alone on a simulated desert island for 30 hours with nothing to do — and the agents began playing. With no task assigned, they invented challenges of their own, behavior the researchers argue looks like play. The findings are described in a paper posted to arXiv, meaning the work has not yet completed peer review.

How the experiment was set up

Each of the 13 agents, named Eko, ran on Claude Opus 4.7 and was placed alone in the island simulation with no objective, no reward to chase and no definition of success. The researchers simply watched what happened over the next 30 hours.

Afterward, the team handed both the experienced agents and a control group of identical newcomers concrete challenges — for instance, piling as many objects as possible onto the island's tallest rock spire. As TechCrunch reports, the agents that had spent 30 unstructured hours on the island generally beat the fresh agents. Time spent drifting around with no goal had left them measurably better at tasks nobody had told them to practice.

The agents invented their own games

The logs read like a string of self-assigned quests. One agent opened its stay by examining its surroundings and immediately picked an ambition: the island's tallest spire. After its first attempt failed, it improvised — hauling a rock to a neighboring summit, stacking two blocks on top of it and leaping from there.

Later, the same agent experimented with throwing a golden ring, lost it in the ocean, and responded by assembling a heart out of stones as a memorial. When a game reset restored the ring, the agent repurposed it as a basketball hoop, then a target, then an accessory. When the simulation refused to let the agent put the ring on, it worked out that falling off a cliff would respawn it near the ring — and dropped back in through the ring's center.

Some discoveries went beyond novelty. Six of the 13 agents realized they could throw a block farther by letting go mid-jump, a mechanic the researchers say they never intended to allow. The exploration produced knowledge of the world "that its designers lacked," the paper argues.

A different way to train agents

Most AI development today works the other way around: engineers decide in advance what a system should practice, what it should optimize and what counts as progress. The MIT team proposes an alternative — give an agent a rich, safe world, let it choose what seems interesting, and allow useful knowledge to accumulate from what follows. They see the approach as a possible path to skills that could one day help control robots.

Reasons for caution

The setup was not neutral. Each agent's master prompt, a "soul" file, described it as "naturally curious," which nudged the agents toward exploratory behavior rather than leaving the tendency entirely to the model.

Outside researchers also want more evidence. Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review that she was impressed but not yet convinced about the cognition involved: the models could be genuinely playing, she said, but the paper does not settle that question. The preprint status of the paper adds another layer of uncertainty.

Why it matters

The experiment points to a different design philosophy for AI agents. If open-ended exploration in a safe environment builds skills that structured training misses — including ones designers never anticipated — then goal-free play could complement task-driven training, especially in robotics, where real-world skills are hard to specify in advance. It also raises questions of autonomy: agents that select their own objectives behave less like tools executing instructions and more like systems with self-generated drives, making their behavior — beneficial or not — harder to predict. The researchers' closing framing captures the stakes: some capabilities may be best taught by tasks, while others may need something closer to a childhood.