Claude agents given no task make games on a virtual island
Agents with 30 hours of play were more able to do set tasks than new copies, but no other group examined the paper.
Claimed, not confirmed
Researchers at MIT put copies of Claude Opus 4.7, each with the name Eko, on a virtual island for 30 hours. No one gave the agents a task. The agents made games and made drawings with blocks. After that, they were more able to do set tasks than new copies. No other group examined the paper.
How it works
Each agent wrote instructions that moved its avatar on the island. It kept text notes in files, which were its long term memory. A soul file told the agent to find new data. When the agent stopped, an instruction told it to continue.
The numbers
The source gives no scores. In one test, the agents had to get as many objects as possible to the highest peak. The agents with play time were more able to do this task than new copies. When researchers removed lines from the memory files, the performance of the agents decreased.
What is not known
Nathan Cloos, an author of the paper, tells that the agent can only copy play. Eunice Yiu of UC Berkeley tells that the paper has no data on the cause of the play. The setup also made the agents do the same work that persons do.
Insight: The memory of an agent holds its skill and its secrets
In this paper, the performance decreased when lines were removed from the memory. A different test, MemCut, shows that agent memory keeps signs of a secret after removal: a reader found it in 39% to 76% of tests. Memory reads as the data that an agent uses and also the data that you must remove. A team that uses agents with memory must test the memory.
Sources
- AI researchers stranded Claude on a virtual desert island. It started to “play.”technologyreview.com
Posted