Core answer
Start here. This is the direct spoken answer to practice first.
Overview
An agent is not the default architecture for every task involving unstructured text.
I avoid an agent when the steps and rules are known, exactness is required, a normal API or search query solves the problem, or the possible actions are too risky to delegate. A single model call or deterministic workflow is easier to test, operate, and explain. The agent must earn its added latency, cost, and failure surface through better task outcomes.