Auto mode replaces per-action prompts with a background safety classifier. The classifier judges each proposed action and only blocks or escalates the risky ones.

What it does

When auto mode is on, Claude keeps working without asking for routine tool calls. A lightweight model reviews each call and compares it against your permission rules and safety baselines. Unambiguously safe calls go through; ambiguous or dangerous ones surface as prompts or denials. This is how you get unattended runs without surrendering review.

When to use it

Long autonomous tasks where constant prompts would stall you out.
Background agents in a sandboxed or containerized environment.
Well-understood repos where you've already tuned permission rules.
Pairing with hooks that log or checkpoint for audit.

Gotchas

Auto mode is a research preview. Behavior can change between releases.
The classifier isn't perfect. Keep protected paths and critical deny rules in place.
Prefer sandboxed environments for true unattended runs - auto mode is not the same as bypass mode.

Official docs: https://code.claude.com/docs/en/permission-modes.md#eliminate-prompts-with-auto-mode