Should your AI assistant ask permission every time?

There's an interesting tension with AI agents.

If it asks you before doing everything, it's basically another productivity tool you have to operate.

If it doesn't ask, you have to trust it enough to act on your behalf.

So where's the useful middle ground?

Personally, I think the answer might be different for different types of actions. Reading, suggesting, drafting and executing probably shouldn't all have the same permission level.

Curious where other people land on this.