Coding agent progress:
The problem I ran into last week was interesting.
In testing, users thought the terminal was freezing because the agents would start work, go silent for several hours, and then hand back a finished product.
Which was, unfortunately, the intended design of the system.
In practice, it turns out agents are boring when you design them like employees.
They just put their little hard hats on, disappear into the mines, and come back later with a pile of finished work.
Turns out that user experience is actually more important than the fact that I built a harness that handles long horizon problem solving on most models larger than 50b parameters.
So I added an operator reassurance lane at the milestone level.
Yes yes, the notifications of progress are the real innovation.
Now the agent spends a few tokens every couple minutes narrating progress, so the human knows the system is alive, oriented, and not quietly dying in a ditch.
This feels like the dumbest thing I’ve ever had to add to an AI system.
But I understand why it was necessary.