Researchers propose a novel framework called "control-data flow separation" to address the instability inherent in prompt optimization within multi-agent Large Language Model (LLM) systems. Currently, prompt optimization often fails because prompts serve dual, entangled roles: generating unstructured, task-relevant content and defining structured execution protocols, such as message routing, output formatting, and termination signals. When optimizers attempt to refine content generation, they frequently inadvertently corrupt these critical protocols, leading to pipeline failures. To mitigate this, the authors decouple these functions by representing execution-critical control as typed, validated program objects, while leaving task-relevant language as the optimizable data flow. This structural separation prevents "prompt drift" from compromising the underlying communication interface. The effectiveness of this approach was evaluated across various domains, including synthetic reasoning, collaborative review generation, and insurance rating workflows. Empirical results demonstrate that this method consistently improves task-specific performance while maintaining a 100% eventual protocol validity rate, ensuring that the multi-agent system's operational integrity remains intact during the optimization process.