What's changing
Anthropic laid it out plainly in its own announcement.
"Starting next week, sessions across Claude web, desktop, and mobile apps will run in auto mode by default on Pro, Max, and Team plans," the company said, adding that "rather than asking you to approve each step, an automated check blocks risky actions.
Switch back to manual anytime." This isn't entirely new ground. Auto mode already became the default inside Claude Code back in August, and now the same idea is spreading to the main apps.
A real-time classifier waves through the safe actions instantly and stops the risky ones, either suggesting a different way to do it or pausing to ask you.
The case for dropping the approve button
The number Anthropic is leaning on is a stark one.
In its testing, the company says the classifier caught about 89% of dangerous actions, against roughly 14% when a person was approving each step manually. The gap isn't really about the machine being clever.
It's about people. After a long session, human attention slips, and anyone who has clicked "approve" fifty times in a row starts clicking it without reading. The careful gatekeeping that manual mode promises quietly falls apart, while a classifier checking every action never gets bored or tired. That's the whole argument for handing the job to software.
Where people still want a hand on it
Not everyone is fully sold. Developers largely welcome the smoother flow, since constant approval prompts are a genuine drag. But some are drawing a line at anything irreversible, deleting data being the obvious one, where an automatic "yes" is a lot more unnerving than a tedious manual one.
Anthropic's answer to that worry is the toggle. Manual mode stays one click away, so anyone doing high-stakes work can switch it on for that session and go back to approving each step.
Why it matters
Underneath a small settings change is a real shift in posture. Anthropic is nudging its paying users toward letting Claude act on its own by default, and backing that nudge with a claim that an automated check protects you better than your own tired attention does.
It fits the direction the whole field has been moving, from AI that asks permission to AI that just gets on with it. The difference here is that Anthropic is making that the starting point rather than an option you opt into, while keeping the manual switch within reach for the moments it matters.