More AI permission still needs a working stop button
Anthropic studied how people use Claude Code. More experienced users often let it act without asking each time, but also stopped it more often. That is an observation in one product. Before granting more freedom, test whether you can pause a run.
What the source reports
Anthropic studied how people use its Claude Code tool. Users with more experience were more likely to approve actions ahead of time. They were also more likely to stop a run while it was in progress. Both facts matter. A permission setting tells the tool what it may do, but a real stop lets a person catch and change a mistake. The study observes behavior in one product; it does not prove that more freedom makes AI safe. Try a low-risk draft with no permission to send anything. Can the named person see the work and stop it before an outside action? If not, keep the task in draft-only mode.
Original source
- Title
- Measuring AI agent autonomy in practice
- Author
- Anthropic
- Date
- Wednesday, February 18, 2026