Graph engineering — representing agentic systems as directed graphs — is a well-established pattern that LangChain has been building with LangGraph for three years. The post explains when to use graphs (structured workflows with predictable paths) versus pure agent loops (open-ended tasks like deep research), and shares key lessons: production agents need cycles not DAGs, loops are just simple graphs, and dynamic transitions via runtime routing (like map-reduce) are essential. What's genuinely new is that nodes can now contain full agents rather than just single LLM calls, enabling practical patterns like embedding a coding agent inside a larger orchestration graph. LangGraph now sees 65M+ downloads per month.
Table of contents
Modeling agents as graphsWhen to represent agents as graphsWhen not to use graphsWhat building LangGraph taught usWhat's actually newThe bigger idea14 Impressions