Great writeup. The point about AI agents needing structure resonates most with me. I've seen too many teams jump to "let the AI figure it out" and end up with unpredictable behavior in production.
The spectrum model you described is important. Not every use case needs a full autonomous agent. Sometimes a structured flow with one AI reasoning step is the right call. The trick is knowing where to draw that line.
On the observability side, I'd add that execution traces are essential once you cross into multi-step workflows. If you can't replay what the agent decided at each step, you're debugging blind when something fails downstream.