The exception path is the product.
Anyone promising agents that run at a hundred percent is selling you a demo. Design for the tail or do not ship.
Lede
Most agent demos show the happy path, because the happy path is the part that demos well. A document arrives, the agent reads it, the right thing happens. It is genuinely impressive and it is also the easy ninety-five percent.
The five percent is where the money is, and it is where projects fail.
What a good exception looks like
An exception is not an error. It is a case the agent correctly declined to decide. The difference matters, because errors get suppressed and exceptions get routed.
A well-formed exception carries four things: what the agent was trying to do, what it found, why it stopped, and what it would have done if forced. That last one is unintuitive and it is the most useful field in the whole record. It converts a human review from "look at this and figure it out" into "confirm or correct" — which is roughly a ten-to-one difference in handling time.
It should also be reversible. If a person overrides the agent, that override is a data point, not just a fix. Ten overrides on the same pattern is a specification change waiting to be written.
The queue nobody watches
The most common failure we see in production agentic systems is not a bad decision. It is a queue with two thousand items in it and no owner.
This happens because exception handling gets designed last, usually by whoever is left at the end of the build, and it gets no service level. Nobody agreed how fast the queue drains or who is accountable when it does not. So it fills, and the workflow it was protecting silently degrades — not because the agent got worse, but because the safety valve stopped being emptied.
Our rule is that the exception queue gets designed before the agent does, with a named owner, a target drain time, and an alert when it is missed. If a client cannot name the person who watches that queue, we do not think the workflow is ready to automate, and we say so.
The uncomfortable version
A useful test: if the agent were switched off tomorrow morning, would anyone notice within an hour?
If the answer is no, the workflow was not important enough to automate. If the answer is yes but nobody knows what they would do about it, then the exception path was never really built — and the ninety-five percent you are so pleased with is resting on nothing.