Your agent has eight known failure types. For each, do you block it inline, measure it offline or send it to a person?

Two decisions hide in this question: where the failure is visible (records, trace or meaning) and when you need to know (before the customer sees it, or after). Most answers collapse them into 'add an LLM judge'. What does a placement that holds up look like?

Updated Nov 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.