Reusable and runtime-adaptive prompt design for dynamic multi-agent LLM systems

Develop prompt design methodologies for large language model (LLM)-based multi-agent systems that are both reusable across tasks and responsive to runtime context changes, particularly in settings where agent roles or task objectives are not predefined, to reduce reliance on handcrafted templates and improve generalization and adaptability.

Background

Prompt design is foundational to how LLM agents interpret instructions and coordinate actions in multi-agent systems. Existing approaches often depend on handcrafted templates or task-specific engineering, which limits generalization across domains and evolving workflows.

The paper highlights a key unresolved challenge: in contexts where agent roles or task goals are not fixed in advance, prompts must be both reusable and able to adapt to runtime context changes. Solving this would enable more flexible and robust multi-agent coordination without heavy manual intervention.