Descripción general del bucle agentic

An agent is just a loop around the Claude Messages API. You send a request, Claude either finishes or asks to run a tool, you run it, and you send the results back. Repeat until Claude is done.

The whole pattern is four steps:

• Request — call the API with your message history.
• Inspect the stop_reason in the response.
• Run tools if Claude asked for them, then add the results to the history.
• Repeat until stop_reason is end_turn.

Master this one loop and you understand the core of every agent.

The Claude API is stateless. The model keeps no memory between calls, so you must send the full message history every single turn.

A request carries these fields:

• model — which Claude model to use.
• max_tokens — the output cap.
• system — the system prompt.
• messages — the entire conversation so far.
• tools — the tools Claude may call.

If you forget to append a turn, Claude simply won't know it happened.

Every response comes back with a stop_reason. This field — not the text — tells you what to do next. The four values are:

• end_turn — Claude finished. The loop is over.
• tool_use — Claude wants a tool run. Execute it and continue.
• max_tokens — output was truncated by the limit.
• stop_sequence — a custom stop sequence was hit.

Your loop is really just a decision based on this one field.

A tool is a function you let Claude call. You describe it with a name, a description, and a JSON Schema for its inputs.

The description is the most important part — Claude reads it to decide when to use the tool. Be clear about its purpose, its inputs, and when it applies.

Claude never runs your tool itself. It only asks for it; your code does the actual work and reports back.

When stop_reason is tool_use, the response content contains a tool_use block. It carries:

• an id — used to match the result back to the request,
• a name — which tool to run,
• an input — the arguments, already parsed as an object.

Claude can ask for several tools in one response. Run them all before continuing.

To continue, you append two things to the history:

1. The assistant's full response.content (so the tool_use block is preserved).
2. A new user message holding a tool_result block.

Each tool_result must include the matching tool_use_id. Then you call the API again — that's one turn of the loop.

Put it together and the agent is a short while loop:

• Call the API with the full history.
• If stop_reason is end_turn, break.
• Otherwise run the requested tools, append the results, and loop.

The model drives the decisions; your code just runs the tools and carries the history.

Here is the rule that separates a robust agent from a fragile one: always terminate on stop_reason, never by reading the text.

Scanning Claude's output for words like "done", "finished", or "complete" is a classic anti-pattern. The model might say "I'm not done yet" or mention the word "done" in a sentence, and your loop breaks at the wrong moment.

The stop_reason field is the one reliable signal. Trust it.

It's wise to add a maximum-iteration counter so a misbehaving loop can't run forever. But understand its role: a cap is a safety net, not the primary way you stop.

The primary stop mechanism is always stop_reason == end_turn. The cap only catches the rare runaway case.

Treating an arbitrary iteration cap as your main stop condition is an anti-pattern — it cuts off legitimate work that just needed one more turn.

In the agentic loop, the model decides what to do: which tool to call, with what inputs, and when to stop. That flexibility is the whole point of an agent.

Reserve hard code for the things you must guarantee — for example, a deterministic check that a refund can't exceed a policy limit, or that an action only runs after identity is verified.

Let the model plan; let your code enforce. Don't hard-code the trajectory, and don't ask a prompt to enforce a critical rule.

Two stop reasons are easy to forget but matter in production:

• max_tokens — the output hit the limit and is truncated. Raise max_tokens or stream, then retry; don't treat partial output as final.
• stop_sequence — a custom stop sequence you configured was matched.

A complete loop branches on all four values. Silently ignoring max_tokens leads to cut-off answers and broken tool calls.

You now know the engine behind every agent:

• Request the API with the full history every turn — the model keeps no state.
• Inspect stop_reason: end_turn (done), tool_use (run tools), max_tokens (truncated), stop_sequence.
• On tool_use, run the tools and append each tool_result with its matching tool_use_id.
• Repeat until end_turn.

Terminate on stop_reason, never by parsing text. Iteration caps are a safety net only. Let the model make decisions; reserve hard code for guarantees.