A prototype can succeed with a prompt, a model, and a happy-path demo. A production AI system has a different contract: it must behave predictably when inputs are incomplete, dependencies fail, users disagree with the model, and business rules change.
That changes the engineering surface. Identity, tool permissions, data boundaries, evaluation datasets, observability, retries, fallbacks, and human escalation become first-class parts of the product rather than infrastructure tickets for later.
The practical lesson is to design the operational boundary at the same time as the model workflow. When an agent can take an action, the system should be able to explain who authorized it, what context it received, what tools it used, and what happened afterward.