IMPROVE: Iterative Model Pipeline Refinement and Optimization Leveraging LLM Agents