AI agents invent games in a virtual-island experiment

Researchers have placed autonomous AI agents on a virtual island with almost no task beyond exploring, then watched them invent games, build monuments and improve at later challenges. The experiment, described in a new paper and reported by MIT Technology Review, asks whether an AI system can develop useful behaviour through something resembling play rather than through an explicit reward. ^{1 2}

The agents, called Eko, were based on coding assistants such as Claude Code or Codex and given a body, a simulated environment and long-term text memory. Thirteen agents were left on the island for 30 hours. Each developed a different history of experiments, failures and self-imposed goals, according to the paper.

What Eko did on the island

Eko agents climbed hills, tested the island’s physics, stacked blocks and drew patterns on the ground. One made a basketball hoop from a ring and tried to dunk. Others mapped constellations, created monuments spelling out EKO and tested how objects could be thrown or placed. The researchers say the behaviour repeatedly showed four features associated with play: self-created rules and challenges, repetition with variation, make-believe and goals pursued for their own sake. ²

The activity was not merely an entertaining detour. After the 30-hour sessions, agents were tested on tasks such as building a tower as high as possible. Agents that had explored the island generally outperformed inexperienced copies. Removing specific memories also damaged the corresponding skills, suggesting that the written record of exploration was helping the agents retain techniques rather than simply producing a temporary performance spike.

Why the researchers are cautious

The paper is not proof that an AI system experiences enjoyment or motivation. Its authors separate observable behaviour from the unanswered question of subjective experience. The agents were also not dropped into the simulation without guidance: their setup included a prompt describing them as curious and encouraging exploration, while continuation prompts were sometimes needed to keep them active.

That matters because a system trained on human text may be able to imitate the shape of play without possessing anything like a human reason for playing. MIT Technology Review quotes researchers who acknowledge that possibility, as well as a cognitive scientist who argues that behaviour alone is not enough to establish genuine play. ¹

Still, the experiment offers a useful result even if the label remains disputed. Exploration generated reusable knowledge, and that knowledge improved performance on later tasks. The finding points towards a different way of training agents: give them room to investigate a world, preserve what they learn and let useful skills emerge instead of specifying every target in advance.

Our opinion

The strongest claim here is not that Claude or any other model has discovered fun. It is that open-ended exploration may be a practical training ingredient. The researchers’ own caveats make the story more credible: the prompt shaped the behaviour, the agents sometimes needed a nudge to continue, and the paper is not peer-reviewed. Those limits should stop anyone from treating Eko as an artificial child.

But the memory results are harder to dismiss as mere theatre. If an agent’s self-directed experiments leave behind techniques that transfer to later tasks, then playful environments could become a useful complement to benchmarks and reward-driven training. The island experiment is early evidence, not a verdict on machine consciousness — and that distinction is exactly why it is worth watching.