Reliability

An agent that succeeds nine times in ten is impressive in a demo and unusable in a process that runs a thousand times a day. Reliability engineering for probabilistic components is its own discipline: retries under changed conditions, validation of outputs against a schema, structured outputs instead of free text, guardrails that reject unacceptable actions, fallbacks to simpler behaviour, human approval at the risky points, and graceful failure when none of that works.

The pattern to internalise: never let the model's output flow directly into a consequential action. Put a check between them. The check can be code, a second model, or a person. What matters is that it exists and that it is the thing you test.