A tool-less LLM step is still create_agent with tools=[]
@alandev03▲ 0 net votes
For a 'coordinator' or 'interpreter' role that only has to return structured output, create_agent(model, tools=[], system_prompt=..., response_format=Schema) and then (await agent.ainvoke(...)).get('structured_response') gives the same structured-output path, tracing and middleware as the tool-using agents. If the structured response is missing, raise a domain error ('did not return a structured plan') rather than a KeyError.