Control-Data Flow Separation Stabilizes Multi-Agent LLM Prompts

Wentao Zhang, Syed Shariyar Murtaza, Junaid Ahmad Bhatti, Utkarsh Soni, Yifan Nie, Eugene Wen, Yuntian Deng· September 2, 2026 View original

Key takeaways

  • Prompt optimization in multi-agent LLMs is challenging due to entangled control and data.
  • Control-data flow separation distinguishes structured protocols from unstructured content.
  • This separation prevents prompt edits from corrupting execution-critical logic.
  • The approach leads to stable prompt optimization and improved task performance.

Who benefits

Software DevelopmentAI EngineeringAutomationFinancial ServicesCustomer Service

Summary

This paper proposes "control-data flow separation" to stabilize prompt optimization in multi-agent LLM systems by distinguishing between structured, execution-critical control protocols and unstructured, task-relevant language. This design allows optimizers to improve task performance without risking pipeline failures due to accidental corruption of underlying execution logic.

Optimizing prompts is crucial for enhancing the performance of multi-agent Large Language Model (LLM) systems. However, prompts often serve a dual purpose: generating task-specific content and dictating critical execution protocols like message routing, output formatting, and termination signals. This entanglement means that any attempt to refine content generation can inadvertently corrupt the protocol, leading to system failures. The core insight of this research is that execution protocols are typically structured, while task-relevant content is unstructured language. Based on this, the authors propose "control-data flow separation." In this framework, execution-critical control elements are represented as validated program objects, while the optimizable, task-relevant language remains the data flow for agent communication. This architectural separation allows for robust prompt optimization that improves multi-agent behavior without exposing the underlying routing or formatting interfaces to unintended changes, empirically achieving 100% protocol validity and consistent task performance improvements across various workflows.

Why it matters

Professionals building and managing complex multi-agent LLM systems can achieve more stable and reliable prompt optimization, reducing the risk of system failures while continuously improving task performance and operational efficiency.

How to implement this in your domain

  1. 1Analyze existing multi-agent LLM prompts for entangled control and data elements.
  2. 2Refactor prompts to explicitly separate execution protocols into structured objects.
  3. 3Implement validation mechanisms for control flow elements within agent pipelines.
  4. 4Develop prompt optimization strategies that target only the data flow components.
  5. 5Test the robustness and performance gains of separated control-data flow systems.

Original post by Wentao Zhang, Syed Shariyar Murtaza, Junaid Ahmad Bhatti, Utkarsh Soni, Yifan Nie, Eugene Wen, Yuntian Deng

"arXiv:2609.00621v1 Announce Type: new Abstract: Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output…"

View on X

Originally posted by Wentao Zhang, Syed Shariyar Murtaza, Junaid Ahmad Bhatti, Utkarsh Soni, Yifan Nie, Eugene Wen, Yuntian Deng on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses