Do AI agents need a better “undo” button more than another confirmation?
A reply on my last thread got me thinking about something I hadn’t really considered before:
Maybe the answer to agent trust isn’t more confirmation. Maybe it’s better recovery.
Right now, the safest UX for an agent often looks like this:
“Do you want me to do this?”
“Yes.”
“Are you sure?”
“Yes.”
“Just checking…”
At some point, we’ve basically built a very cautious coworker who needs emotional reassurance before creating a reminder. 😅
The obvious problem is that confirmations interrupt the exact flow agents are supposed to protect.
But removing them completely doesn’t feel right either.
So I’ve been wondering if the better pattern for low-risk actions is:
Act first. Make the result obvious. Make undo effortless.
For example, if an agent creates a reminder, renames a file, opens an app, or moves something into a folder, I probably don’t need a confirmation dialog every time.
I’d rather see:
Done — reminder created for 3 PM.
Undo
That feels very different from giving an agent unlimited autonomy.
The user is still in control — control just happens after the action instead of before it.
I’m starting to think there are three different UX patterns here:
Low-risk + reversible: act immediately, show what happened, offer undo.
Medium-impact: act or prepare the change, but keep it visible and easy to edit.
External / irreversible: ask first.
What interests me is that “reversibility” might matter just as much as risk.
Deleting a draft that can be restored may actually feel safer than sending a perfectly harmless message that can’t be unsent.
And maybe that’s part of why some agents still feel awkward today:
we’re designing autonomy around permission, when users may actually care more about recoverability.
I’m curious how other builders think about this.
Would you rather have an agent ask before every action, or act more freely if you knew you could undo almost anything instantly?
And where does “act + undo” stop feeling comfortable for you?
Replies