Researchers have introduced a framework called control-data flow separation to address a critical challenge in multi-agent LLM systems: prompt optimization can inadvertently break the underlying execution protocols. The work, presented in a paper on Hugging Face, highlights that prompts often serve two entangled roles—generating task-relevant content and specifying execution-critical protocols such as message routing, output formatting, and termination signals.
The key insight is that these roles have different representations: execution protocols are typically structured, while task-relevant content is expressed in unstructured language. The proposed framework represents execution-critical control as typed, validated program objects, while keeping task-relevant language as the optimizable data flow for agent communication. This design prevents prompt drift from affecting routing or formatting interfaces.
Empirical evaluations across synthetic reasoning, collaborative review generation, and insurance rating workflows demonstrate that the framework achieves 100% eventual protocol validity while consistently improving task performance. This suggests that separating control from data can stabilize multi-agent systems during optimization.
The paper also lists related works, including approaches like LocalLSTC for GUI agents and reasoning-driven search across prompts and programs, indicating active research in this area.