Anatomia de uma Requisição de API

Every call to Claude is one request to the Messages API. As an architect, you don't memorize syntax — you reason about six fields that shape the whole interaction: model, max_tokens, system, messages, tools, and tool_choice.

Get these right and everything downstream — agents, tool loops, structured output — falls into place. This lesson walks through each field and the decision it represents.

The model field selects which Claude does the work. It's a single string, and the choice is a real architectural trade-off: capability versus latency versus cost.

• Opus — most capable, best for long-horizon agentic and hard reasoning.
• Sonnet — strong balance of speed and intelligence.
• Haiku — fastest and cheapest for simple, high-volume tasks.

You can change the model per request, so route easy tasks to a cheaper model and hard ones to a stronger one.

max_tokens is a hard cap on how many tokens Claude may generate in this response. The model is not told this number — it's an enforced ceiling, not a hint.

If generation hits the cap, the response is cut off and stop_reason comes back as "max_tokens". That means the answer is truncated, not complete. Set it high enough to finish the job; for very long outputs, stream so you don't hit request timeouts.

The system prompt sets Claude's role, tone, and standing rules — the instructions that apply to the whole conversation rather than to one user turn.

Put durable behavior here: "You are a support agent. Verify the customer's identity before any account action." Keep it stable across requests — a frozen system prompt also caches well, which lowers cost and latency on repeated calls.

This is the field architects get wrong most often. The Messages API is stateless: the model keeps NO memory between requests. On every turn you resend the full conversation history — every prior user and assistant turn, plus any tool results.

If you only send the latest user message, Claude has amnesia. Conversation state lives in your application; you replay it each call. Messages alternate roles, and the first message must be user.

The tools field is a list of actions Claude may call — each with a name, an input_schema (JSON Schema), and most importantly a description.

The description is the primary way Claude decides which tool to use — not the name. A good description states the tool's purpose, its return values, input formats with examples, and when NOT to use it. Vague or overlapping descriptions cause misrouting. Keep the set tight: about 4–5 tools per agent is optimal; piling on 18+ degrades selection reliability.

tool_choice controls whether and how Claude calls a tool:

• {"type": "auto"} — Claude decides whether to reply with text or call a tool (the default).
• {"type": "any"} — Claude MUST call some tool. This is how you guarantee structured output.
• {"type": "tool", "name": "X"} — force one specific tool.

Use any or a forced tool when you need a machine-readable result every time; use auto for open conversation where a plain answer is sometimes correct.

Every response carries a stop_reason that tells you what to do next. Architects branch on it — they never parse the text looking for words like "done".

• "end_turn" — Claude finished naturally; the turn is complete.
• "tool_use" — Claude wants a tool run; execute it, append the result, continue.
• "max_tokens" — output was truncated by the ceiling.
• "stop_sequence" — a configured stop string was hit.

This single field is the control signal for the entire agentic loop.

Tools, messages, and stop_reason combine into the core pattern: request → inspect stop_reason → if tool_use, run the tools and append results to history → repeat until end_turn.

Because the API is stateless, you append the assistant's tool request AND the tool result back into messages before the next call. Termination is driven by stop_reason — decisions are model-driven. Any iteration cap you add is a safety net, never the primary stop mechanism.

Statelessness isn't a limitation to work around — it's the design that makes Claude predictable and scalable. Because the model holds no hidden state, the request is the complete truth: same six fields plus the same history produce the same behavior.

This is why context management is its own discipline. As history grows you trim verbose tool output to the relevant fields, summarize old turns, and keep critical transactional facts (IDs, amounts, dates) verbatim in a dedicated block — because the model only knows what you put back in messages.

A production request is rarely just a model and a prompt. A support-agent turn combines all six fields: a routed model, a safe max_tokens, a rules-bearing system, the full messages history, a tight tools set, and a tool_choice that matches the task.

Read these six fields off any request and you can predict exactly how it will behave — that's the architect's lens.

You now have the architect's mental model of a Claude request:

• model — capability vs. cost vs. latency; route per task.
• max_tokens — enforced output ceiling; stop_reason: "max_tokens" means truncated.
• system — durable persona and rules; keep it stable.
• messages — the FULL history, resent every turn, because the model holds no state.
• tools — descriptions drive selection; keep to ~4–5 well-scoped tools.
• tool_choice — auto / any / forced; use any to guarantee structured output.

And the loop that ties them together: drive it on stop_reason (end_turn vs tool_use), never by parsing text. Master these and the rest of the certification builds on solid ground.