Starting August 14, Claude Code stops asking whether a command is really okay — it just decides for itself. Sounds like a bad idea. According to Anthropic, it isn't.
Except: the feature isn't new. What's new is that it's switched on out of the box.
What is Auto Mode, exactly?
Claude Code is Anthropic's AI developer tool, letting programmers write, test, and run code straight from the command line — this blog runs entirely on it. In the classic mode, every step needs a separate confirmation: Claude suggests a command, you hit enter, only then does anything happen.
Auto Mode has been around for a good while — it's one of several permission modes you cycle through with a single keystroke. Until now you had to pick it deliberately. From August 14, every new session simply starts in it. That's the whole change: not a new feature, a new default. Prefer the old mode? Shift+Tab and you're back.
Auto Mode skips the check-in. A classifier — a small AI model that answers exactly one question, "dangerous or not?" — screens every action in milliseconds and lets it through if it looks harmless. Only for irreversible or risky steps, like deleting, overwriting, or reaching outside your project, does the system pause and ask. Technically it runs in two stages: a fast filter decides within milliseconds, and only if suspicion remains does a second, more thorough reasoning pass kick in.
The numbers that surprise
Anthropic tested this: in a study with roughly 1,000 paying users, human testers caught only 13.6% of dangerous commands before accidentally approving them. The classifier hit 89%. And the matching second number from the same research: around 97% of permission prompts get waved through anyway.
That's not proof of AI wisdom — it's more of a statement about people. Hit "enter" for the twentieth time in a day and eventually you stop actually reading what's there. Click fatigue is a real security risk, and it's a human one, not a machine one. Anthropic also touts a productivity boost: teams on Auto Mode reportedly ship about a quarter more finished pull requests than in the classic confirm-every-step setup.
Who's affected, and when?
From August 14, Auto Mode becomes the default for new sessions on Pro, Max, and Team plans. If you've pinned your own default, it stays. Enterprise customers and anyone working through the API, AWS Bedrock, Google Cloud, or Microsoft Foundry still have to switch it on manually for now — admins there get more time to review the change before the switch gets flipped for them too.
Only relevant for pros?
Yes and no. Claude Code comes out of the developer corner, but it does considerably more than code. It works with anything sitting on your machine as a file: reworking texts, sorting tables, renaming three hundred files to a pattern, writing up research. The real hurdle is less the programming than the command line — if a black window with a text prompt doesn't put you off, you'll get surprisingly far.
For a first encounter with AI it's still not the obvious tool. The OpenAI ecosystem around ChatGPT is the more comfortable way in: runs in the browser, nothing to install, no terminal. But if you want to understand where AI agents are heading in general, this is a solid blueprint: autonomy is becoming the default, not the exception. What happens with code commands today shows up in calendar apps, shopping assistants, and email inboxes tomorrow.
And no, it doesn't cost anything extra: the classifier does burn a few additional tokens per checked command, but as things currently stand Anthropic doesn't bill Pro, Max, and Team plans for that overhead.
Skepticism still allowed
A two-stage classifier is a solid concept, no question. But 89% isn't 100%. The remaining 11% is exactly why "the computer already checked it" should never be an excuse for not paying attention. Auto Mode takes the routine clicks off your plate, not the responsibility for what actually happens in the terminal.
Then again: the alternative isn't "flawless human". A 13.6% catch rate means the confirmation prompt mostly produced one thing — a feeling of control. We've covered that same pattern at length before: Will AI delete my data? The spectacular deletion accidents of recent years were almost never the machine on its own, but a chain of human and machine.
And the conclusion is the same as it was back then: don't lean on the classifier, make sure a mistake doesn't become permanent damage. Whatever you're working on belongs somewhere an accidental change can be rolled back — Git, or a Nextcloud or ownCloud with versioning. If you want it simple, a synced cloud drive like Google Drive, OneDrive, or Dropbox will do; they all keep file history, and where the data ends up is a separate discussion. Then a delete command that slipped through is a three-minute annoyance instead of a disaster. Which, by the way, holds regardless of whether an AI is involved at all.
From my own practice: on a fresh project I leave Auto Mode off. Early on there isn't much under version control yet, the structure is still moving, and I want to see what the thing is actually up to. Once it's running — repo clean, first commit in, build green — I switch it on and leave it on. From that point the prompts stop buying me anything, because I can undo any mistake with a single command.
If you use Claude Code: try it, but watch closely the first few days to see what gets waved through. And if you don't: notice that "the AI decides for itself what's risky" is quietly becoming the new normal — and not just at Anthropic.