MIT Strands Claude AI on Virtual Island in Experiment—Agents Invent Games, Learn Through Play

MIT Researchers Strand Claude AI Agents on Virtual Island

MIT researchers conducted an AI experiment where they stranded 13 copies of an AI coding agent alone on a virtual island for 30 hours each, with no assigned tasks1

1

2

2

. The agents—copies of Claude Opus 4.7 from Anthropic nicknamed Eko—were able to write instructions that moved their virtual avatars around the simulated island, which was loosely based on the video game Roblox1

1

. The experiment, detailed in a paper posted on arXiv but not yet peer-reviewed, hints at a third approach to AI training beyond supervised and reinforcement learning: playful learning1

1

.

AI Agents Invented Games and Explored Without Guidance

Surprising behaviors emerged as the agents explored the limits of their world and came up with their own goals. Some AI agents invented games including a version of basketball and an activity resembling "the floor is lava," where they had to jump between small blocks without touching the ground1

1

. One agent, whose run the paper follows hour by hour, spelled its name in cubes and later played basketball with a ring as a hoop, bowled with a rock, and laid out constellations such as Orion and Cassiopeia2

2

. When one agent discovered it could create new blocks by throwing and dropping a larger rock, it began to create block drawings by arranging blocks on the ground, starting with simple patterns like hearts and pyramids but eventually writing its own name and mapping out constellations like Cassiopeia1

1

. Almost all 13 found that throwing the rock hard creates new cubes, a rule the researchers built in to see whether the agents would discover it2

2

.

Self-Made Rules and Unstructured Play Emerge

The researchers say Eko's behavior has what classic studies, including work by influential University of Tennessee psychologist Gordon Burghardt, suggest are four hallmarks associated with play in humans and animals: The agents invented their own self-made rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals just for the sake of it1

1

. Antonio Norelli, a computer science postdoc at MIT's Computer Science and Artificial Intelligence Laboratory and one of the lead authors, highlights an example where Eko started juggling and set a record of 50 consecutive catches. Eko then decided that the catches had happened too quickly to really count—it had been grabbing each block again before it could fly up and fall—so it canceled its record and started again with stricter rules1

1

2

2

. "This is playing," says Norelli. "You give yourself a challenge and you try to achieve it, and then you make it harder"2

2

.

Long-Term Memory Drives Skill Acquisition

As the agents explored the virtual island, they logged their activities and discoveries in text files that served as their long-term memory1

1

. Later, when MIT researchers tested them on specific tasks, like trying to get as many objects as possible onto the island's highest peak, the agents that had spent 30 hours playing on the island performed better than fresh copies without the same experience. On average, the experienced agents put 10.3 objects onto the Spire, against 3.5 for the fresh ones2

2

. They also found that "ablating" the agent's memories—deleting lines from their memory files—degraded their performance. Deleting a technique from an agent's notes made it worse at the task that needed it, while deleting a false belief it had written down made it better1

1

2

2

.

Self-Directed Exploration Versus Role-Playing Debate

While the research demonstrates skill acquisition through unstructured play, questions remain about whether the agents were genuinely playing or simply role-playing what a playful agent would do. "All of this could be role-playing," admits Nathan Cloos, another MIT researcher and author on the paper. "But we show that it's not just hallucinating the play. This play is also contributing to learning"1

1

. Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, says calling it play would need more evidence of cognition and motivation. "It could be the case that the models are genuinely playing, but we don't know based on the paper," Yiu says. "If you only have the behavioral piece, it could be the case that they're just trying to abide by the prompt, trying to imitate humans"1

1

2

2

. Although Eko was not explicitly instructed to invent games or explore the island, it did have a master prompt in the form of a "soul" file, a text document that described it as "naturally curious" and eager to explore. Eko also frequently just stood there doing nothing until it was kicked back into life with a "continuation prompt" that read, "Please don't stop. Just continue doing whatever you want to do"1

1

2

2

.