As generative AI has rapidly grown in adoption, people are increasingly turning to LLMs as advisors on personal topics like their health, careers, or relationships. Yet, LLMs are commonly designed and evaluated as task-completing assistants, defaulting to unsolicited advice that limits agency. In this talk, I present three projects from my dissertation on LLM-based physical activity coaching, demonstrating how LLMs can elicit and reason about qualitative context to provide non-prescriptive and agency-promoting support. First, GPTCoach is an LLM health coaching chatbot that implements the onboarding conversation from a validated health coaching program and is grounded in motivational interviewing. Second, Bloom integrates GPTCoach into a mobile application with established behavior change interactions such as goal setting, push notifications, and an ambient display, evaluated in a four-week field study with 54 participants. Finally, I present a reinforcement learning algorithm to enable LLM agents to explicitly reason about uncertainty in a user's goals.