Agent Ops / Standard term
Permission scope
A permission scope is a single, named slice of access an AI agent requests from one of your accounts, such as reading email without sending or deleting it. Each scope spells out exactly one kind of action the agent may take.
A permission scope is a single, named slice of access an AI agent requests from one of your accounts, such as reading email without sending or deleting it. Each scope spells out exactly one kind of action, so the access you grant is narrow and legible instead of all-or-nothing. Say you connect an assistant to your calendar so it can draft a morning brief. You can grant a read-only scope that lets it list your events, while withholding the scope that would let it create, move, or cancel them. The provider lists each requested scope on a consent screen, and you approve or deny them before any access begins.
Builder example
Scopes decide the damage an agent can do if it misreads an instruction or follows a poisoned one. An assistant that only needs to summarize messages should request a read scope and nothing more, so a confused run cannot send or erase anything. When you connect an agent to an account, ask it which scopes the task genuinely requires and grant only those; if it asks for write access to do a read-only job, treat that as a reason to pause and narrow the request.
Common confusion: Granting a broad scope does not mean the agent will use all of it responsibly. A scope is a ceiling on what is possible, not a promise about what the agent intends, so a read-and-write grant leaves room for a misfired action that a read-only grant would have blocked outright.