Separating Control From Data in Multi-Agent Prompts Stops Optimizers From Corrupting the Routing Protocol
Prompts in multi-agent systems serve two entangled roles, generating task content and specifying execution-critical protocol such as message routing, output formatting, and termination signals that the surrounding code depends on, so an edit meant to improve content can silently break the pipeline. The authors' observation is that these roles have different natural representations, protocols being structured and task content unstructured, so they represent execution-critical control as typed validated program objects while leaving task language as the optimizable data flow. Across synthetic reasoning, collaborative review generation, and insurance rating workflows this reaches 100% eventual protocol validity while still improving task performance.
↳ Follow the thread