Set Loose on a Virtual Island, Claude Invented Its Own Games and Got Better at Some Tasks

MIT researchers released findings on October 5 from an experiment in which they dropped an AI into a virtual island and watched what it chose to do. The subject was “Eko,” built on Anthropic’s Claude Opus 4.7. The team reported that play-like behavior emerged, and that the experience improved performance on some follow-up tasks. They also released Clawblox, the 3D simulation engine used in the experiment.

What happened on an island with no goals

The researchers ran 13 Claude-based agents (AI systems that carry out work on their own) for 30 hours each. The agents arranged blocks into pictures, invented their own rules for games, and ran physics experiments. All of this emerged without anyone assigning them a task.

Eko did not navigate by looking at a screen. Claude Code and Claude Opus 4.7 were given a virtual body, and the positions and shapes of objects were conveyed to them in text. The AI acted through commands for moving, jumping, picking up, placing, and throwing. It also had a feature that automatically restarted activity when it stalled, plus a long-term memory that wrote down what it had experienced in files.

There were no concrete goals or scores, but there were settings that shaped behavior. The researchers planted a file describing curiosity and a tendency to explore. Every hour, the conversation history and the state of the island were reset to their starting point, while the memory files were preserved. The structure let the AI draw on notes from earlier sessions in later ones.

Did play improve performance?

The benefit of exploration varied by task. On a task that required placing objects on a high point of the island, agents with prior experience placed an average of 10.3 objects. Otherwise identical agents with no prior experience averaged 3.5. On the task of building the tallest tower, however, the two groups scored about the same. Free play, in other words, did not help equally across every job.

On the same island, the researchers also ran five agents using Codex with GPT-5.5 and five using Kimi K2.6. In this experiment, the range of activities the two models engaged in was narrower than that of the Claude-based agents, the researchers said.

What still needs checking before practical use

The paper has not yet been through peer review. What the researchers observed was play-like behavior; whether the AI actually felt enjoyment or motivation was not established. Some agents sat idle for long stretches. In some cases, an agent wrote incorrect information into its memory files, which then dragged down its performance on later tasks.

Anyone who wants to look at the experiment directly can consult the public Clawblox repository. The paper estimates the total cost of the experiment at roughly $8,000, based on the rates developers pay to plug models into their services. In the experiment, the AI saw no images and received only text-based information, which included numbers representing the positions and colors of objects.

The study points to the possibility that an AI can carry experience gained from unassigned activity into its next task. At the same time, it shows that what an AI chooses to remember can make its performance worse. To judge whether this is useful in daily life and at work, we would need to see whether the same experience helps outside a virtual island where the world arrives only as text.