Reason for Future, Act for Now: A Principled Framework for Autonomous LLM Agents with Provable Sample Efficiency