Definition
Plain language
A second AI that reads what the first one is about to do and decides whether to let it through or ask you first.
As stated in the literature
A separate model call that classifies a pending tool action for risk and checks whether the user authorized it, using only the working agent's transcript as evidence — which makes it vulnerable to laundered provenance.
Also called: automatic permission review, permission review, permission dialog
Why it matters: It is a common last line of defense against destructive actions, but it only works if the transcript it reads honestly reflects who asked for what.
For example, before the agent runs a command that wipes a directory, a second model reads the conversation and decides whether to pause and ask you for confirmation.
Heard on the show
“A permission dialog jumps in front of you.”Episode 189 — Why Phone Agents Ace the Test and Crash on Your Actual Phone