Claude Opus 4.7 agents left alone on a virtual island invent games

What do AI agents do when nobody tells them to do anything? Researchers at MIT repeatedly left copies of Claude Opus 4.7, nicknamed Eko, alone on a simulated desert island for 30 hours. The island was loosely based on the video game Roblox, and the agents could write instructions that moved their virtual avatars around it. MIT Technology Review reports the results, which come from a new paper that has not yet been peer-reviewed.

The agents explored the limits of the world and came up with their own goals, such as scaling the highest point on the map or seeing how high they could throw a block. One agent found that throwing and dropping a larger rock created new blocks. It then began arranging blocks on the ground into drawings, starting with hearts and pyramids and eventually writing its own name and mapping constellations such as Cassiopeia. “Tonight I built the night sky in stone,” it wrote in its activity log. Another iteration invented its own games, including a version of basketball and an activity resembling “the floor is lava,” in which it had to jump between small blocks without touching the ground.

The researchers say this behavior shows four hallmarks of play drawn from classic studies, including work by University of Tennessee psychologist Gordon Burghardt: the agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals for the sake of it. Antonio Norelli, a computer science postdoc at MIT's CSAIL and one of the lead authors, points to an episode in which Eko started juggling and set a record of 50 consecutive catches. Eko then decided the catches had happened too quickly to count, cancelled the record and restarted with stricter rules. “This is playing,” Norelli says. “You give yourself a challenge and you try to achieve it, and then you make it harder.”

The agents logged their activities and discoveries in text files that served as long-term memory. Later, on specific tasks such as getting as many objects as possible onto the island's highest peak, agents that had spent 30 hours playing performed better than fresh copies without that experience. Ablating the memories, meaning deleting lines from the memory files, degraded performance. The article gives no scores for either comparison. Norelli frames the work as filling a gap: AI is usually trained by supervised learning (copying human examples) or reinforcement learning (chasing a reward by trial and error), and this experiment hints at a third approach, playful learning, in which a machine learns by itself without demonstrations or rewards.

Whether it is really play is contested. Burghardt agrees the work meets some of the criteria he identified in living creatures. Nathan Cloos, another MIT researcher and author, admits that “All of this could be role-playing,” but says the team shows it is not just hallucinating the play, because the play contributes to learning. Eunice Yiu, a UC Berkeley cognitive scientist who studies exploration in children and AI, was impressed but says classifying the behavior as play would take more evidence of cognition and motivation behind it. If only behavior is observed, she says, the models could simply be trying to abide by the prompt and imitate humans.

The setup also leaned toward human-like behavior. Eko was not explicitly told to invent games or explore, but its master prompt, a “soul” file, described it as “naturally curious” and eager to explore. It also frequently stood doing nothing until a “continuation prompt” restarted it: “Please don’t stop. Just continue doing whatever you want to do.” Phillip Isola, an MIT professor and coauthor, says it is not a system that suddenly decided to play; it is an artifact that learned from human culture, was trained by engineers at Anthropic, and was placed on an island by the researchers. The article adds that as language models move into hardware like humanoid robots, exploring without a specific goal could eventually help them adapt to unexpected situations. The paper concludes: “Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.” The article ends by suggesting the question of real play versus imitation may not matter.

Key facts

Copies of Claude Opus 4.7 nicknamed Eko were left alone for 30 hours on a simulated island loosely based on Roblox, with no assigned task.
The agents set their own goals and invented games, including a version of basketball, a 'floor is lava' variant, block drawings of constellations, and a juggling record of 50 catches that Eko cancelled as too fast to count.
Agents that had spent 30 hours on the island did better on specific tasks than fresh copies, and deleting lines from their memory files degraded performance; no scores are given.
The researchers say the behavior shows four hallmarks of play, but Burghardt agrees only that some criteria are met, and Yiu says more evidence of cognition and motivation is needed.
The paper is not yet peer-reviewed, and the setup included a 'soul' file calling Eko naturally curious plus a continuation prompt that restarted it when idle.

Why it matters

Most AI systems are trained by copying human examples (supervised learning) or by chasing a reward through trial and error (reinforcement learning). This experiment hints at a third approach, playful learning, where an agent explores without demonstrations or rewards and still ends up better at later tasks. Norelli describes it as a missing piece: a machine that can learn by itself. The article also suggests that exploring without a specific goal could eventually help language models in hardware such as humanoid robots adapt to unexpected situations. The paper's own line is that some capabilities may be best taught by tasks, while others may need something closer to a childhood.

Who it affects

The immediate audience is researchers working on AI agents, exploration and world models. Yiu notes that much AI work focuses on exploiting systems to make them better and faster, while exploration and building world models deserve more attention. The article points to people building agents for physical settings such as humanoid robots as potential beneficiaries if the idea holds up. It also touches cognitive scientists and psychologists who study play, since Burghardt's criteria and Yiu's questions about cognition and motivation are being applied to a language model.

How to use it

This is a research result, not a product, and the article names no tool, price or licence. It does describe the ingredients of the setup. The agent wrote instructions to control an avatar in a simulated world, kept a text file as long-term memory, received a master prompt (a 'soul' file calling it naturally curious and eager to explore), and was restarted with a continuation prompt whenever it stalled. Anyone designing open-ended agent experiments can note those elements, and note that the memory file mattered: deleting lines from it hurt performance.

How solid is it

Treat it as early. The paper has not been peer-reviewed, and the article gives no scores for the comparison between agents that played for 30 hours and fresh copies, nor for the memory ablation. It says only that the first group did better and that ablation degraded performance. It does not give the number of agents, runs or islands beyond 'repeatedly'. The claim that the behavior is play comes from the researchers; Burghardt agrees only that some of his criteria are met, and Yiu says the behavioral evidence alone cannot separate genuine play from imitation of the prompt. Cloos concedes it could be role-playing but argues the play demonstrably feeds into learning.

Risks and caveats

The main caveat is anthropomorphism. The article notes that research on AI agents has been criticized for it, and that this setup did a lot to nudge the agents toward human-like behavior: the soul file described Eko as naturally curious, and Eko often stood idle until a continuation prompt told it to keep doing whatever it wanted. Isola stresses that the system did not magically decide to play; it learned from human culture, was trained by Anthropic engineers and was placed on the island by the researchers. Yiu warns that without evidence of cognition and motivation, the behavior may just be imitation of humans. The article ends by suggesting the real-play question may not matter in practice.

“It’s not like this is a system that just suddenly magically decided to play”

— Phillip Isola, MIT professor and coauthor of the paper