Why LangGraph

An LCEL chain is a DAG you drew at design time — data flows forward through a fixed sequence of steps and stops. That model breaks down the moment an application needs a loop (retry until a check passes), a branch chosen at runtime (the model decides which tool, if any), or a pause (wait for a human before continuing). LangGraph is the orchestration layer built for exactly that: a graph of nodes and edges, with a state object that flows between them and can cycle back on itself.

Key idea

LangGraph doesn't replace LCEL — a graph node is often just a compiled LCEL chain. LangGraph adds the loop, the branch, and the durable state around chains that individually stay linear.

create_agent (see Agent Concepts) is itself a thin, pre-built LangGraph graph. This section covers what you build once the pre-built agent isn't shaped like your problem: custom state, custom routing, checkpointed persistence, and multi-agent handoff.

See also

• Agent Concepts — the chain-vs-agent decision this section builds on.
• State and Nodes — the core StateGraph building blocks.