The agentic loop — the foundation of every Agent SDK app

The agentic loop is the four-step cycle that turns a stateless LLM call into a system that takes actions in the world.

The loop:

1. Send the conversation history (plus the tools available) to Claude.
2. Look at the response's stop_reason field.
3. If stop_reason === "tool_use" — execute the requested tools and append their results to history.
4. If stop_reason === "end_turn" — the loop is done. Return the final response.

That's it. Every Agent SDK app — single-agent or multi-agent — is built on this.

---

stop_reason is the single most important field on the response. It is the canonical signal for loop control.

Value	What it means	Loop action
"tool_use"	Claude wants to call one or more tools before continuing	Execute the tools, append tool_result blocks, loop again
"end_turn"	Claude has finished its turn	Exit the loop

You'll see exam questions suggesting other ways to detect completion — checking the assistant's text content, counting iterations, sentiment analysis. All wrong. The right answer is always: inspect stop_reason.

---

Why "model-driven" matters. A naive engineer might write:

if (userMessage.includes("refund")) callTool("process_refund");

This is not an agentic system. The decision has been pulled out of Claude and into your code. Every new edge case becomes a new if branch.

Agentic systems push decision-making into Claude:

1. Give it good tool descriptions
2. Give it relevant context
3. Trust it to pick the next action

Your code's job is the loop infrastructure — sending requests, executing tools, handling errors. Not the decisions themselves.

---

Anti-patterns to memorise:

• ❌ Parsing assistant text to decide when to stop
• ❌ Using an iteration cap as your primary stopping mechanism (it's a safety belt, not a brake)
• ❌ Treating "any text content" as a completion signal — Claude can emit text and tool calls in the same response
• ❌ Pre-classifying input before Claude sees it (routing logic that runs before the model)

If a question shows you a system doing any of these, the right answer is almost always "fix it by leaning on stop_reason and the model's judgment."