Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development