MoMHa is a system that optimizes large language model (LLM) harnesses by considering multiple objectives: accuracy, safety, and token cost. It uses an agentic proposer to search for optimal harness design and outperforms alternative approaches in various domains.