Try to draw your agent's steps on a whiteboard. If you can draw them all, you have a workflow, that's a finding worth acting on.
Run the decision prompt with honest answers, particularly about how much the path really varies.
Answer question 3 seriously: could a workflow handle the common cases and an agent only the awkward ones? This hybrid is underused and often correct.
Check whether you can capture full traces. If you can't, that's a blocker, not a detail.
Write down the lowest autonomy level that still delivers most of the value. Ship that one first.