Research questionHow can in-car LLM agents respond consistently to incomplete requests they cannot safely fulfill?An agent may succeed on a request occasionally while failing on repeated attempts or proceeding when it should clarify, defer, or refuse. In-car assistance makes this difficult because user intent is often incomplete and responses may depend on stateful dialogue and tool use.