Your loop around the support agent ends the conversation whenever Claude's response contains a text block. Logs show many sessions end right after Claude writes "Let me pull up that order for you" in the same response as a lookup_order tool_use block, so the lookup never runs and the customer gets no answer. How should the loop decide when to continue?

Answer: D — Inspect stop_reason: when it is "tool_use", execute the requested tools, append their results to the conversation, and call Claude again; end the loop only on "end_turn".

stop_reason is the loop's control signal: "tool_use" means Claude wants tool results before it continues, and "end_turn" means it has finished (task 1.1). Treating text content as a completion signal is a named anti-pattern, and so is parsing the text for phrases (a), which is brittle natural-language control flow. Banning text next to tool calls (b) relies on the model to keep the loop correct and still checks the wrong signal. A fixed iteration count (c) is an arbitrary cap used as the main stopping mechanism: it cuts off cases that need more steps and wastes calls on cases that need fewer.