The rapid rise of agentic AI in mental health exposes a misalignment between conversational fluency and user safety. We present a multidimensional framework spanning nine evaluation domains—from clinical validity to relational risk and regulatory compliance—designed to assess agentic systems beyond surface metrics. Applying it to eight real-world tools reveals that the most popular agents often lack basic safeguards, while clinically validated solutions remain underutilized. These findings highlight the need for structured oversight and public awareness. Our framework offers a foundation for safer deployment, informed regulation, and design practices grounded in mental health ethics.