Configurable multi-agent framework for scalable and realistic testing of llm-based agents