Skip to content
How does it decide to ask versus act?
It classifies each action by three questions — is it reversible, is it external (does it leave your accounts), and is it sensitive. Reversible, internal, non-sensitive things are handled quietly; sensitive, irreversible, or un-permissioned things are proposed and wait for your approval.
If it's routine and within a standing permission you've granted, it acts under that permission and reports back. If it's uncertain whether something is sensitive, it surfaces it as a decision rather than guessing. The safe default always wins.
See also: What is configured autonomy? · When does the assistant act on its own?