MASPO: Joint Prompt Optimization for LLM-based Multi-Agent Systems

Researchers introduce MASPO, a framework that automatically optimizes prompts across multi-agent LLM systems by evaluating how well each agent's outputs enable downstream success rather than in isolation. The approach uses evolutionary beam search to navigate prompt spaces and achieves 2.9% average accuracy improvements over existing methods across six diverse tasks.

MASPO addresses a fundamental challenge in LLM-based multi-agent systems: the misalignment between optimizing individual agent performance and achieving overall system goals. Traditional prompt optimization treats agents independently, missing critical interdependencies where one agent's output quality directly impacts another's input quality and downstream success. This research bridges that gap through a joint evaluation mechanism that contextualizes each prompt's value within the broader system architecture.

The framework emerges from growing recognition that as LLM applications become more complex, simple prompt engineering no longer suffices. Multi-agent orchestration requires understanding how agents interact—a problem intensifying as organizations deploy increasingly sophisticated AI pipelines for reasoning, planning, and task execution. MASPO's evolutionary beam search provides computational efficiency crucial for exploring the exponential prompt space without exhaustive evaluation.

For developers building production AI systems, this work offers practical methodology to improve multi-agent reliability without ground-truth labels—a significant constraint in real-world deployment. The 2.9% accuracy improvement, while mathematically modest, compounds across complex task chains where cascading errors typically multiply. Organizations using LLM systems for customer service, code generation, or business process automation benefit from better agent coordination at the prompt level.

The public code release signals research transparency valuable for the community. Moving forward, attention should focus on scalability to larger agent networks, generalization across different LLM architectures, and integration with emerging frameworks handling agentic workflows. Understanding how MASPO performs with heterogeneous agent types and dynamic task environments will determine its real-world applicability.

• →MASPO optimizes multi-agent LLM prompts jointly rather than independently, addressing misalignment between local agent objectives and system-wide goals.
• →Joint evaluation mechanism assesses prompts by their capacity to enable downstream agent success without requiring ground-truth labels.
• →Evolutionary beam search efficiently navigates high-dimensional prompt spaces across distributed agent systems.
• →Framework demonstrates 2.9% average accuracy improvement over state-of-the-art methods across six diverse collaborative tasks.
• →Open-source release accelerates adoption and research in multi-agent prompt optimization methodologies.