We already have this concept. It’s called user accounts.
Your Gmail account vs my Gmail account. Your macOS account vs my macOS account.
Yes, I can spam you from my Gmail. Yes, I can use sudo on my Mac and damage your account. But the impact is by default limited.
The answer is to just treat assistants as a different user profile, use the same sharing mechanisms already developed (calendar sharing, etc), and call it a day.
Isn't this what the parent is saying ?
That's punting the problem in the same way SELinux did. Agent loops are useful precisely because they're zero config.
Problem: I want to accomplish work securely.
Solution: Put granular permission controls at every interface.
New problem: Defining each rule at all those boundaries.
There's a reason zero trust style approaches won out in general purpose systems: it turns out defining a perfect set of secure permissions for an undefined future task is impossible to do efficiently.