MIT study: Claude agents left on a virtual island started to play

What the research found

MIT researchers ran 13 Claude Code (Opus 4.7) agents on a virtual island for 30 hours each with no task; the agents showed play-like behaviour and experienced agents did better on later tasks, but the arXiv paper is not yet peer reviewed.

Claims and how well the evidence supports them

1. 13 Claude Code agents running Claude Opus 4.7 were left alone on a virtual island for 30 hours each 97%
2. The paper is posted on arXiv and has not yet been peer reviewed 99%
3. Experienced island agents placed 10.3 objects on the Spire versus 3.5 for fresh agents 99%
4. Deleting a false belief from agent notes improved task performance 98%

Evidence

• thenextweb.com https://thenextweb.com/news/mit-claude-agents-virtual-island-play

Actionability rated none.