Three Years of Graph Engineering with LangGraph

Graph engineering is an approach that represents agent systems as graphs, emerging as the latest concept following prompt engineering and context engineering. LangGraph is a framework that implements this approach, and is currently downloaded more than 65 million times a month, used by both startups and large enterprises.

In a graph, nodes perform the actual work, while edges determine the next step. Edges can be deterministic, or they can branch conditionally based on node results or state. This makes it possible to clearly separate the parts that require an LLM's judgment from the parts that code must forcibly control.

However, graphs are not always the right answer. For tasks like deep research, where it's difficult to define the path in advance, the agent harness approach is more suitable. The key is to use graphs where deterministic control is needed, and open loops where full autonomy is needed.

In practice, the two approaches are often mixed. For example, you can define fixed steps like classify→retrieve→synthesize as a graph, while allowing subagents to operate autonomously within each step. LangGraph supports this kind of multi-agent architecture, where each node itself can be a complete agent with its own internal loop.