Trial and Error: Exploration-Based Trajectory Optimization for LLM Agents