There is one moment in agentic products that I don’t think any of us have gotten right yet, my own explorations included. It’s the moment the agent hits a wall and has to stop and ask for help.
We pour real hours into the prompt box, the loading state and the little checkmark when the task is done. Then the one moment where the whole thing can quietly fall apart, the interruption, gets whatever the generic error component happened to look like that quarter.
I don’t think that’s laziness. It is hard to design for a failure you can’t fully predict. But this is where trust survives or doesn’t. Not the happy path, this one.
Silence reads as broken. A task that simply stops, with no explanation, doesn’t look like an agent being careful. It looks like it crashed. We can’t blame users for that reading, because we never gave them a reason to read it any other way.
I’m working on a clearer answer, but I don’t have a tidy framework to hand you today. What I have is a growing conviction that this is the moment worth obsessing over next, more than the parts we usually obsess over.
If you hand an asynchronous task to an agent in your product today, could you draw what happens between the delegation and the outcome, including the part where it stops and asks?