When should an AI agent act immediately — and when should it ask first?

I’ve been thinking about a product-design problem that seems simple at first, but gets complicated very quickly:

When should an AI agent just do the thing — and when should it stop and ask for confirmation?

The more actions an agent can take, the more useful it becomes.

But the more autonomy you give it, the more important trust becomes.

I’ve started thinking about this less as an “AI capability” question and more as a combination of risk, reversibility, and external impact.

For example:

Low-risk + reversible
Open an app. Search Maps. Create a reminder.
These feel like actions an agent should probably just do.

Medium-risk + editable
Create a calendar event. Draft a message. Rename or move a file.
Maybe the agent should act, but make the result easy to review or undo.

Higher-risk + external
Send a message. Delete something. Submit a form. Make a purchase.
These feel much more like “ask first” territory.

What I find interesting is that too much confirmation can make an agent feel useless.

If I have to approve every tiny step, I’m basically doing the work myself again.

But too little confirmation creates the opposite problem:

the agent may save clicks while adding anxiety.

So the UX question I keep coming back to is:

How do you preserve the feeling of “just get it done” without making users feel like they’ve lost control?

One framework I’m currently testing is:

Act immediately when an action is low-risk and easy to reverse.
Act + show when it’s editable and recoverable.
Ask first when it creates an external or hard-to-reverse consequence.

I’m curious how other people building agentic products think about this.

Do you use a similar risk / reversibility framework, or have you found a better rule for deciding when an agent should act vs. ask?