- Given a curious persona, copies of Claude Opus 4.7 spent 30 hours on a virtual island, invented games, and later did better than fresh copies on tasks.
- The agents kept memory in text files. Deleting lines from those files lowered performance. The paper is not yet peer-reviewed, and its setup nudged the agents toward play.
- Leaders deploying agents should ask where agent memory is stored, who can edit it, and what the agent's founding instructions say.
The value may sit in a text file
Most talk about AI agents focuses on the model. A new study points somewhere else. In it, a plain text file of the agent's own notes appears to be what made the agent better.
MIT Technology Review reports on a paper from MIT researchers. The paper has not yet been peer-reviewed. In the experiment, the team ran copies of Claude Opus 4.7, which they called Eko, through 30 hours alone in a simulated world. The world took loose inspiration from the video game Roblox. The agents moved their avatars by writing instructions.
Eko was not told to invent games or explore. Its instructions did describe it as curious and told it to keep going. The lesson for a leader is less about whether machines can play. It is that an agent's experience can live in a file that people can read, edit or delete. That makes the file something to govern.
What the agents did on the island
The agents set their own goals. They tried to climb the highest point on the map. They tested how high they could throw a block.
One agent found that throwing and dropping a larger rock created new blocks. It then arranged blocks into drawings. It began with hearts and pyramids and moved on to its own name and the constellation Cassiopeia. "Tonight I built the night sky in stone," it wrote in its log.
Another version invented a form of basketball and a game like "the floor is lava." In a third example, Eko juggled and reached 50 consecutive catches. It then judged the catches too quick to count, cancelled the record and restarted with stricter rules.
How the setup worked
The agents wrote their activities and discoveries into text files. These files served as long-term memory. A language model does not otherwise carry experience from one session to the next.
Later, the researchers set specific tasks, such as getting as many objects as possible onto the island's highest peak. Agents with the island history performed better than newly started copies.
The team also deleted lines from the memory files, a step called ablation. Performance dropped. That suggests the notes carried useful knowledge and were not just a diary.
How much the setup shaped the result
MIT Technology Review's report raises doubts about the design. Research on AI agents has been criticised for treating them like humans, and this setup nudged them that way.
Eko had a master prompt called a "soul" file. It described the agent as "naturally curious" and eager to explore. Eko also frequently stood still until a "continuation prompt" restarted it. That message asked it not to stop and to carry on with whatever it wished.
Phillip Isola, an MIT professor and coauthor, said this is not a system that "suddenly magically decided to play." He described it as an artifact trained by Anthropic's engineers and placed on an island by the researchers.
Whether it was true play is also disputed. Nathan Cloos, a paper author, admits it "could be role-playing." He says the play still contributed to learning. Eunice Yiu studies exploration in children and AI at UC Berkeley. She was impressed, but she says the paper cannot settle the question. In her view, calling it play needs evidence about the thinking and motivation behind the behaviour, and the paper offers only observation. Without that, she said, the agents could simply be doing what the prompt asks and mimicking people.
What this shows for organisations
Set the debate over play aside. Two points from the study matter for anyone deploying agents. First, a founding document describing the agent as curious was part of the setup. The report says it nudged the agents toward human-like behaviour. Second, the agent's notes appear to have shaped how well it performed.
This is our interpretation, not a finding of the paper. An agent that accumulates notes accumulates an asset. That asset can be lost, damaged or altered, and the study shows that removing lines changes results. The open question is whether your teams treat such files as records with owners, backups and change history.
The study also comes from a game-like island, not a business system. It does not show how agents behave with company data.
Questions to put to your team
Where does each of our agents store its memory, and who can edit it? Is there a record of changes to the founding instructions that describe how the agent should behave? If we delete or reset an agent's notes, do we know what capability we lose?
Ask also whether agents are tested against a fresh copy with no history. That is the comparison the MIT team used, and it separates what the model brings from what experience adds.
The paper itself says some capabilities may be best taught by tasks, while "others may need something closer to a childhood." For a buyer, the practical point is simpler. If an agent learns on the job, someone has to look after its notes.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: MIT Technology Review.
