Lessons Learned from Evaluation of LLM based Multi-agents in Safer Therapy Recommendation