CCA-F: Claude Certified Architect — Foundations

In the Claude API's tool-use (agentic loop) pattern, each API response includes a stop_reason field indicating why the model stopped generating. When Claude wants to invoke a tool, the response has stop_reason: "tool_use". The orchestrating application should check this field after every API call: as long as stop_reason is "tool_use", it executes the requested tool(s), sends the tool results back as a new user message, and calls the API again. Once stop_reason changes to "end_turn" (or another terminal value such as "max_tokens" or "stop_sequence"), Claude has finished its turn and the loop exits. Checking for the presence of a text block is unreliable because Claude can emit text alongside a tool_use block in the same response (e.g., explaining its next action) without that meaning it's done. Setting tool_choice to "none" is a way to prevent tool calls on a specific request, not a dynamic stopping signal, and tracking a maximum call count is merely a safety cap, not the primary control mechanism.