tool_choice: automático / qualquer / forçado
Deixe o modelo escolher, force uma ferramenta ou fixe uma pelo nome.
tool_choice: automático / qualquer / forçado é uma aula grátis de Claude Architect no CoddyKit. Esta é a aula 2 de 4. Você pode ler a aula completa abaixo gratuitamente — depois pratica ao vivo no navegador com um editor de código integrado e um tutor de IA 24/7. Faz parte do caminho de aprendizado de Claude Architect, e seu progresso é sincronizado entre a web e o app CoddyKit. O curso de Claude Architect inclui 4 aulas no total.
Partes desta aula ainda não foram traduzidas e aparecem em inglês.
Who Decides: Model or You?
When you give Claude tools, one question shapes the whole interaction: who decides whether a tool runs?
The tool_choice field on your API request answers it. You can let the model decide, force it to call some tool, or pin one specific tool by name.
Getting this right is core to Tool Allocation: the wrong setting produces chatty text when you needed structured data, or a forced tool call when the model should have just answered.
The Three Modes
There are three values for tool_choice:
auto— the model freely picks: emit text, or call a tool.any— the model must call some tool (its choice which), so the turn returns a tool call, never free text.{"type":"tool","name":"X"}— force one specific named tool.
Each maps to a different intent: flexible reasoning, guaranteed structured output, or a hard-pinned action.
request = {
"model": "claude-sonnet-4-5",
"max_tokens": 1024,
"tools": tools,
"tool_choice": {"type": "auto"}, # or "any", or a named tool
"messages": messages,
}auto: Let the Model Choose
auto is the default mindset for agentic loops. The model inspects the conversation and decides on its own whether a tool is needed or a plain text reply suffices.
Use it when the path is open-ended: a support agent that might answer a question directly, or might need to look up an order first.
With auto, the stop reason tells you what happened: tool_use means run the tool and continue the loop; end_turn means the model answered in text and is done.
resp = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
tools=tools,
tool_choice={"type": "auto"},
messages=messages,
)
if resp.stop_reason == "tool_use":
# run the requested tool(s), append results, loop again
...
elif resp.stop_reason == "end_turn":
# model replied with text; turn complete
...any: Guarantee a Tool Call
any forces the model to call some tool every turn — it cannot reply with free-form prose. It still chooses which tool, but a tool call is guaranteed.
This is the classic lever for guaranteed structured output: if your only tool is a schema-shaped record_result, then any means every response comes back as validated JSON arguments, never an unparseable paragraph.
classify_tool = {
"name": "record_sentiment",
"description": "Record the sentiment of a customer message.",
"input_schema": {
"type": "object",
"properties": {
"sentiment": {"type": "string",
"enum": ["positive", "neutral", "negative"]}
},
"required": ["sentiment"],
},
}
resp = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=512,
tools=[classify_tool],
tool_choice={"type": "any"}, # must call a tool -> structured output
messages=messages,
)Forced: Pin One Tool by Name
The most specific mode forces exactly one tool: {"type":"tool","name":"X"}. The model has no choice about which tool — only how to fill its arguments.
Use it when the action is already decided and you just need the model to extract the parameters. Example: you know this turn must produce an extraction record, so you pin extract_invoice and let Claude fill the fields.
resp = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
tools=[extract_invoice_tool],
tool_choice={"type": "tool", "name": "extract_invoice"},
messages=[{"role": "user", "content": invoice_text}],
)
# resp.content[0] is a tool_use block with the extracted fields
fields = resp.content[0].inputany vs forced: The Subtle Line
Both any and forced guarantee a tool call. The difference is choice:
any— model still selects which tool from the set. Good when several tools are valid and you want structured output but flexible routing.- forced
{"type":"tool","name":"X"}— no routing decision at all; tool X runs.
Rule of thumb: if you've already made the decision in code, force the tool. If the model should still decide which structured action fits, use any.
Structured Output Is the Big Win
Pairing a JSON-Schema tool with any or a forced tool is how architects get reliable structured output. The schema eliminates syntax errors and enforces required fields — no fragile regex on prose.
One schema rule matters here: mark a field required only if it is always present. Never require a field that may be absent — the model will fabricate a value to satisfy the schema.
extract_invoice_tool = {
"name": "extract_invoice",
"description": "Extract structured fields from an invoice document.",
"input_schema": {
"type": "object",
"properties": {
"invoice_number": {"type": "string"},
"total": {"type": "number"},
"due_date": {"type": "string"}, # may be absent -> NOT required
},
# only fields that are ALWAYS present belong here
"required": ["invoice_number", "total"],
},
}auto Keeps the Agentic Loop Alive
In a multi-step agent, auto is usually right because the loop is model-driven. You send the full history each turn, read stop_reason, run any requested tools, append results, and repeat — until the model emits end_turn.
If you forced a tool on every turn, the model could never signal completion with text, and your loop would have no natural stop. Reserve forcing for single, decided actions.
while True:
resp = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=1024,
tools=tools,
tool_choice={"type": "auto"},
messages=messages,
)
messages.append({"role": "assistant", "content": resp.content})
if resp.stop_reason == "tool_use":
results = run_tools(resp.content)
messages.append({"role": "user", "content": results})
continue
break # end_turn -> terminate on stop_reason, never on textDon't Confuse Choice with Enforcement
tool_choice controls whether a tool is called — it does not enforce business rules. Forcing process_refund does not guarantee the refund is allowed.
Critical guarantees (identity verified, refund under the policy cap) belong in deterministic preconditions and hooks, which are 100% reliable, not in tool_choice or prompt text (~90% probabilistic).
Think of tool_choice as routing, and hooks/preconditions as guardrails.
Description Still Drives Selection
Even with auto or any, the model's tool pick depends on tool descriptions, not names. A good description states purpose, return values, input formats with examples, edge cases, and applicability boundaries.
Overlapping or vague descriptions cause misrouting that no tool_choice value can fix. Keep about 4-5 well-scoped tools per agent; 18+ degrades selection reliability.
tool_choice sets the policy; descriptions make the routing accurate.
A Practical Decision Guide
Pick by intent:
- auto — open-ended agent turns; the model may answer in text or act. Default for the agentic loop.
- any — you need structured output but several tools could apply; model routes among them.
- forced (name) — the action is already decided; you only need argument extraction (classification, single-shot extraction).
If you've made the decision in code, force it. If the model should reason about whether and which, use auto or any.
# Single-shot classification: decision already made -> force it
resp = client.messages.create(
model="claude-sonnet-4-5",
max_tokens=256,
tools=[record_sentiment_tool],
tool_choice={"type": "tool", "name": "record_sentiment"},
messages=[{"role": "user", "content": review_text}],
)
sentiment = resp.content[0].input["sentiment"]Quick Check
Choose the best tool_choice for the scenario below.
Recap: Choose, Guarantee, or Pin
Key takeaways:
auto— model picks text or a tool; the default for model-driven agentic loops that must be able to signalend_turn.any— must call some tool; the lever for guaranteed structured output with flexible routing.{"type":"tool","name":"X"}— pins one tool; use when the action is decided and you only need argument extraction.tool_choiceis routing, not enforcement — put critical guarantees in hooks/preconditions.- Accurate selection still depends on strong tool descriptions and 4-5 scoped tools per agent.
Perguntas Frequentes
A aula “tool_choice: automático / qualquer / forçado” é grátis?
Sim — o texto completo de “tool_choice: automático / qualquer / forçado” é grátis para ler aqui na web. Para praticá-la interativamente (um editor de código integrado e um tutor de IA 24/7) e desbloquear o restante do curso de Claude Architect, atualize para CoddyKit PRO. O curso de Claude Architect inclui 4 aulas no total.
O que vou aprender em “tool_choice: automático / qualquer / forçado”?
Deixe o modelo escolher, force uma ferramenta ou fixe uma pelo nome. Você pratica Claude Architect com código prático que executa diretamente no navegador, e um tutor de IA 24/7 responde suas dúvidas enquanto trabalha na aula.
Preciso ter experiência prévia para começar Claude Architect?
Nenhuma experiência prévia é necessária. Claude Architect no CoddyKit é estruturado para alunos iniciantes até avançados, então você pode começar aqui ou desde o início e aprender no seu ritmo. Esta é a aula 2 de 4.
Quanto tempo leva a aula “tool_choice: automático / qualquer / forçado”?
A maioria das aulas CoddyKit leva cerca de 5–10 minutos. Cada uma é compacta e interativa, então você faz progresso constante e retoma exatamente de onde parou entre web e app.
Posso escrever e executar código nesta aula de Claude Architect?
Sim. Cada aula de Claude Architect inclui um editor de código integrado, então você escreve e executa código real direto no navegador e recebe feedback de IA instantaneamente — nenhuma configuração local necessária.
Todas as aulas deste curso
- Quantas Ferramentas por Agente
- tool_choice: automático / qualquer / forçado
- Ferramentas Integradas do Claude Code
- Padrão de Investigação Incremental