Definition
Plain language
Letting an assistant run its own commands without a person clicking "yes" each time.
As stated in the literature
A harness setting in which tool calls are cleared by an automated policy or classifier rather than an interactive human confirmation, enabling long unattended agent sessions.
Also called: auto-approve, auto approval
Why it matters: Auto-approval is what makes hours-long unattended agent work possible, but it also relocates the security boundary from a human's judgment to a policy someone had to write in advance.
For example, a build agent might be allowed to run any read-only command like listing files or checking test results without pausing for a human, while anything that touches the network still needs a person.
Heard on the show
“If you've got auto-approval on, in a repo with credentials and network access, that monitor is the thing saying no.”Episode 269 — How a Forged Transcript Got Model Weights Past a Safety Monitor