The one reasoning step

One step in this process needs reasoning: is this company a good fit for what we sell, and what should the rep know before quoting? A rep used to spend ten minutes on the web for this. Now an Agent node does it, and the way it is built is the whole lesson.

A fixed shape. The agent must answer with five named fields: fit, reasons, suggested products, questions to ask, and evidence. That is the output schema. Output validation checks the answer against it before anything downstream can read it, and sends it back up to twice with the failure explained. Everything after this step reads the five fields by name.

One tool, and a budget. The agent has web search and nothing else. Its limits allow four iterations, four tool calls and eight thousand tokens; the whole step has a sixty-second timeout and each search fifteen seconds. Reach a limit and the step ends cleanly. The product catalogue is in the instructions; a larger one would be bound as knowledge.

Never invent. The instructions say that if the web gives nothing, the agent must say so and leave fit as unknown. In the recorded run the search tool failed, and the agent did exactly that: fit unknown, evidence stating the search failed, one sensible question to ask the prospect. A rep reading that knows what happened. A model allowed to guess would have written a confident paragraph about a hotel it had never seen.