Traditional end-to-end LLM calls often suffer from hallucinations or logical leaps when facing multi-step, multi-constraint complex tasks.