Since August 14, Pro, Max and Team subscribers have started their sessions with the agent authorized to act unless an irreversible or destructive action is detected.
Anthropic reversed Claude Code's default setting. Since August 14, 2026, the new sessions of the Pro, Max and Team formulas have opened in automatic mode: the agent starts its actions without asking the user, instead of asking for a validation at each tool call. Enterprise accounts, API and cloud platforms must follow within one month.
The guard is no longer the human but a classifier, which examines each tool call and interrupts execution when it detects irreversible, destructive or outgoing action. Anthropic puts forward a counter-intuitive argument to justify the switch: on a deliberately dangerous set of controls, the automatic filter intercepted 89%, where only 13.6% of the human rereaders could find it.
Click fatigue
The underlying reasoning is the known one of fatigue in the face of alerts: a user who is asked to approve forty consecutive actions ends up approving the forty-first without reading it. Users who have explicitly chosen another default setting received a one-time invitation to switch, which they remain free to refuse.
Read the full article on the original source
Go to the source →