After a number of projects, the reasons an agent never reaches production repeat themselves. None of them has anything to do with which model was chosen.
1. The agent doesn't have the data
It is asked to prepare a report, but half the information lives in a shared folder nobody wants to grant access to. The result is an agent that writes well and gets the numbers wrong. Access has to be solved before anything is built — with the minimum scope necessary.
2. Nobody defined the boundary
If it isn't written down what the agent may do without approval, one of two bad things happens: either it is allowed too much and an incident appears, or it is allowed nothing and the saving never arrives. The boundary is a business decision, not a technical one.
3. It is measured against a feeling
"It feels faster" does not survive the first budget cut. Without a baseline taken before you start — minutes per task, errors, rework — there is no way to defend the project or to know whether to continue.
4. The exception got automated
It is tempting to attack the hardest case to prove the technology. It is the classic mistake: the rare case eats the budget and returns little time. Do the frequent and boring first; the exception afterwards, once it is isolated.
5. Nobody stayed in charge
An agent is software: formats change, rules change, models change. If nobody in-house can adjust it when the project ends, it degrades on its own. That is why the last phase is always handover and training, not a closing presentation.