Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs

Multi-agent LLM systems often rely on prompts that serve two entangled roles: generating task-relevant content (e.g., summaries, critiques) and specifying execution-critical protocols (e.g., routing, formatting, termination). When prompt optimization methods are applied to these systems, they often inadvertently modify the protocol instructions, leading to system failures, misrouting, or crashes. This paper asks: How can we optimize agent prompts while ensuring the underlying execution protocol remains stable and robust?

The authors propose control-data flow separation, a design principle that decouples the two roles of agent outputs. In this framework, every agent output is split into two channels: a structured, typed control channel consumed by the program controller, and an unstructured data channel containing task-relevant content. The control channel is governed by a schema that is validated at runtime and kept in a frozen prompt slot, making it inaccessible to the prompt optimizer. The optimizer is only permitted to modify the data channel, ensuring that improvements to agent behavior cannot corrupt the execution interface.

Across four settings—including synthetic reasoning, collaborative review generation, and insurance underwriting—the authors demonstrate that their framework achieves 100% eventual protocol validity. In contrast, naive prompt optimization frequently causes system collapse on complex multi-agent tasks. By isolating the control interface, the authors show that prompt optimization can consistently improve task performance without the risk of breaking the agent pipeline. The framework is also shown to be robust across different LLM families, including OpenAI, Anthropic, and Google models.

This work provides a practical solution to a critical failure mode in autonomous agent systems. By applying established software engineering principles—specifically the separation of concerns—to prompt-based systems, the authors enable the use of powerful automated optimization techniques in complex, multi-agent environments. This makes it possible to build systems that are both highly performant and reliable enough for sensitive, real-world applications.