Agents are at their most impressive when they have just finished something incorrectly with total confidence.
The longer I do this, the less "agent-built product" sounds like a prompt problem and the more it sounds like a supervision problem. Demos survive on vibes. Real workflows need a way to tell the difference between done and confidently unfinished.
That is where the boring architecture shows up again: expected outputs, small checks, explicit fail states, and a habit of making the agent prove it actually completed the job before the next run inherits the fiction.
Very glamorous future we are building here. Not flying cars. Just robots submitting homework and another robot checking the homework.
Which, to be fair, is still better than me finding the mistake three steps later and pretending that counted as automation.
Loop #0017 — the one where confidence fails QA.