LLM-based agents have emerged as innovative tools capable of performing complex tasks through iterative planning and action, making significant advances in understanding and addressing user needs. However, their effectiveness is limited in specialized domains such as mental health diagnosis, where they underperform compared to general applications. Current approaches to integrating diagnostic capabilities into LLM rely on rare and highly sensitive mental health datasets, which are difficult to obtain. Furthermore, these methods fail to mimic clinicians' pre-questioning skills, lack multi-session conversational comprehension, and struggle to align their results with expert clinical reasoning. To address these gaps, we propose DSM5AgentFlow, the first LLM-based agent workflow designed to autonomously generate DSM-5 Level-1 diagnostic questionnaires. By simulating therapist-client conversations using specific client profiles, this framework provides transparent, step-by-step disease predictions, generating explainable and reliable results. This workflow serves as a complementary tool for mental health diagnosis that adheres to ethical and legal standards. Through comprehensive experiments, we evaluate leading LLMs across three key dimensions: conversational realism, diagnostic accuracy, and explainability. Our dataset and implementation are fully open source.
