Survey on Agent Workflow Optimization: From Static Templates to Dynamic Runtime Graphs

The IBM research team has compiled a list of papers in the field of LLM agent workflow optimization, along with the survey paper 'From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents' (arXiv:2603.22386). This repository systematically summarizes two major research directions—static optimization and dynamic adaptation. The core argument is: Agent workflows are shifting from static predefined templates to dynamically adaptive runtime graphs. This survey provides a comprehensive academic map for this emerging field and is worth in-depth reference.

LLM agents are evolving from simple question-answering systems to autonomous agents that perform complex multi-step tasks. However, traditional manually designed static workflow templates have obvious limitations:

• Lack of adaptability: Fixed templates struggle to handle task diversity and dynamic changes;
• Suboptimal performance: Manual design is often not the optimal solution;
• High maintenance cost: New scenarios require redesign;
• Poor scalability: Manual design becomes infeasible as the number of agents increases. Therefore, workflow optimization has become a key challenge in organizing agent collaboration.

Static optimization determines the workflow structure before execution and is divided into four categories:

1. Offline template search: Use evolutionary algorithms, reinforcement learning, etc., to search for optimal templates in the design space (e.g., AFlow, A²Flow, SEW);
2. Fixed-frame node-level optimization: Keep the topology unchanged and optimize individual node configurations (e.g., DSPy, CAPO, Optima);
3. Joint optimization of structure and configuration: Optimize both topology and node parameters simultaneously (e.g., Maestro, Multi-Agent Design);
4. Verifiability applications: Ensure workflows meet safety constraints (e.g., MermaidFlow, VFlow).

Dynamic optimization allows adjusting workflows during execution and is divided into three categories:

1. Selection and pruning: Dynamically select relevant agents and prune irrelevant parts (e.g., Cut the Crap, AgentDropout, MasRouter);
2. Build-then-execute: Dynamically generate workflows before task execution (e.g., AutoFlow, WorkflowLLM, G-Designer);
3. In-execution editing: Modify the structure while executing to handle unexpected situations (e.g., DyFlow, AgentConductor, MetaGen).

Combing through the papers reveals four major trends:

1. From manual design to automatic evolution: Use evolutionary algorithms, meta-learning, etc., to automatically optimize workflow structures;
2. From general-purpose to task-specific: Task-specific optimization (difficulty awareness, domain feature adaptation);
3. From efficiency to robustness: Focus on safety, verifiability, and error recovery (e.g., RobustFlow);
4. From single-agent to multi-agent collaboration: Communication topology and collaboration protocol design become core.

Implications for practitioners:

1. Choose appropriate strategies: Use static optimization for fixed tasks, dynamic optimization for dynamic scenarios, and verifiable methods for high-reliability scenarios;
2. Adopt incrementally: Gradually transition from node-level optimization to structural optimization and runtime adaptation;
3. Focus on evaluation metrics: In addition to task success rate, consider efficiency (token usage, latency), robustness, and interpretability.