Dynamic Path Navigation for Motion Agents with LLM Reasoning