Most useful agent work starts as operations work. Who owns the output? What counts as a failure? Which source of truth wins? What should be logged? Who gets interrupted when confidence is low?
The model is one part of that system, not the system itself. Good automation has boring edges: retry rules, approval points, clear naming, small runbooks, and enough logging to debug a bad day.
That is why I treat agent projects as workflow design first and implementation second. The code matters, but the operating model is what makes it safe to use every week.