If you use Claude Code regularly, you know the drill: you kick off a bigger task, lean back — and 30 seconds later, the first permission prompt pops up. Can I write this file? Can I run this bash command? Five times, ten times, twenty times per session. You end up babysitting the tool instead of doing something else.
What Auto Mode Changes
As of March 24, there’s a new option: auto mode. Instead of approving every single action, Claude decides on its own whether a step is safe enough to execute. Before each tool call, a classifier reviews the planned action and only blocks what actually looks risky: mass file deletion, data exfiltration, suspicious code. Everything else goes through.
This might sound like the existing --dangerously-skip-permissions flag. It’s not. That flag skips all checks entirely. Auto mode has a real safety net: when the classifier flags an action as risky, it doesn’t just stop — it actively redirects Claude to find a different approach.
Who This Is For
The typical use case: you kick off a larger refactoring task and want to step away for half an hour. Or you let Claude write tests overnight. Until now, you either had to sit there watching or disable safety entirely. Auto mode is the middle ground.
The caveat: Anthropic explicitly recommends using auto mode in isolated environments — containers, separate git branches, that sort of thing. It’s a research preview, not a finished feature. Available first for Claude Teams users, with Enterprise and API access following in the coming days.
My Take
This is one of those quiet but important steps. Not the feature that makes headlines — but the one that concretely changes how developers work day to day. The balance between autonomy and safety is handled pretty cleverly here. And as Simon Willison pointed out: auto mode makes --dangerously-skip-permissions truly obsolete for the first time.
Sources: