RAG & Agent System Design: the practice test

If the right passage never came back, you have a retrieval bug: fix chunking, indexing, or the query, and no prompt work will help. If it came back and the model ignored or contradicted it, you have a generation bug with a completely different toolbox. One log lookup separates them, which is why seasoned people ask this before theorizing. Teams that skip it reflexively rewrite prompts to fix retrieval failures, the single most common wasted week in RAG debugging.

An agent buys you dynamic control flow: the model decides what to do next. That is valuable exactly when the path is unpredictable, and pure cost when it is not. Manual Q&A is retrieve-then-answer, a two-step pipeline with no decisions to delegate; wrapping it in an agent loop adds latency, token spend, non-determinism, and a new class of looping failures, for nothing. The senior instinct interviewers look for is reaching for the simplest structure that fits the task, and defending it.