In multi‑agent large language model systems, prompts serve two intertwined roles: they generate task‑relevant content and they encode execution‑critical protocols such as message routing, output formatting, and termination signals. The protocol part is usually structured and type‑checked, while the content part is free‑form natural language. Optimizing a prompt without separating these roles can corrupt the protocol and cause the whole agent pipeline to fail. Observing this mismatch, we propose control‑data flow separation: execution control is represented as typed, validated program objects that remain outside the prompt’s mutable space, whereas task‑related language stays as the optimizable data flow for agent communication. Optimizers therefore modify only the language component, preserving routing and formatting interfaces. Experiments on synthetic reasoning, collaborative review generation, and insurance rating workflows show that protocol validity stays at 100% across all scenarios while task performance consistently improves.
Review