From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents

A team of nine researchers, led by Ling Yue, has published a comprehensive survey on arXiv that provides a much-needed framework for understanding and optimizing workflows for LLM-based agents. The paper, 'From Static Templates to Dynamic Runtime Graphs: A Survey of Workflow Optimization for LLM Agents,' introduces the concept of Agentic Computation Graphs (ACGs) to describe systems that interleave LLM calls, tool use, code execution, and memory. The core contribution is a novel organizational lens that categorizes methods based on when a workflow's structure—its components, dependencies, and information flow—is determined. This cleanly separates static methods, which use a fixed, reusable scaffold, from dynamic methods that can select, generate, or revise the workflow for a specific task at runtime.

Beyond this primary categorization, the survey organizes prior work along two other critical dimensions: what part of the workflow is being optimized and which evaluation signals (like task metrics or execution traces) guide that optimization. The authors also make a crucial distinction between reusable workflow templates, the specific graphs deployed for a given run, and the final execution traces. To push the field forward, they advocate for a 'structure-aware evaluation' perspective. This means complementing traditional task success metrics with analysis of graph-level properties, execution cost, robustness, and how the structure varies across different inputs. The goal is to establish clearer vocabulary, a unified framework for comparing methods, and more reproducible evaluation standards for the rapidly evolving domain of LLM agent optimization.