When agents go wrong: orchestrate first
When agents go wrong, they go wrong confidently.
That's why agentic AI isn't always the answer, even when the model could do the whole job. Whether AI is executing a tool or suggesting an answer, a few bad responses and the user stops trusting it. That's when I've found orchestration becomes essential. Orchestration owns the steps and their order in code, and the model does the parts that need judgment inside each one.
An invoice is a good example. Read it, match it to the purchase order, check the totals, flag what looks off. Those steps don't change, so I don't let a model choose them. Each step has a narrow job and an output I can check, and only the ones that need judgment go to a model. When something goes wrong, I know which step did it.
Matching invoice lines to a purchase order works the same way. The code pulls the candidate PO lines for that vendor and enforces the hard rules, like same currency. The model pairs the lines and gives each pair a confidence and a plain reason. The person sees the suggestion and why, and accepts it or doesn't, before anything gets paid.
Agents do have their place, in work where the next step depends on what was just found. An invoice that matches nothing is a good case. But even there, the agent runs inside code. Its tools are limited, its output is checked, and when it isn't sure, the case goes to a person.
I expect agents to take on more as their track record builds, one kind of case at a time.
Where do you let an agent run free, and where have you gone back to a fixed workflow?