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

The survey published on March 25, 2026, in ArXiv AI, focuses on the evolution of workflow optimization within large language model (LLM) systems.

It emphasizes the transition from static templates to dynamic runtime graphs, which allow for more flexible and efficient execution of workflows that integrate LLM calls.

This shift reflects the increasing popularity of LLM-based systems, which are being utilized for a variety of tasks through executable workflows that combine LLM interactions with information retrieval and tool usage.