The demo optimizes for possibility.
Production optimizes for repeatability. The difference is mostly invisible until a workflow is used by real people with incomplete data, ambiguous requests and strict permissions.
Evaluate the unit of work.
Do not score only final prose. Measure retrieval quality, tool choice, action correctness, latency and escalation behavior as separate stages.
“A reliable agent is a state machine with language in the loop.”
Human review is not failure.
The best systems make it obvious when human judgment is needed and preserve enough context that the handoff is fast.