Course Content

What is LangGraph, and when would you choose it over LangChain Chains or Agents?

Each step down was forced by one new requirement: the refund bot only needed a graph once finance demanded a durable human approval.

What you need to know

The three layers in 2026

The LangChain ecosystem was reorganised around the 1.0 releases. It helps to see it as three layers:

Layer	What it is	Use it for
LCEL chains (`prompt \| model \| parser`)	A fixed, forward-only pipeline	One pass, no loop, no pause
`langchain.agents.create_agent`	A ready-made tool-calling agent, extended with middleware	The standard "model calls tools until done" loop
LangGraph `StateGraph`	The runtime underneath both	Custom branches, loops, approvals, multi-agent flows

Older names you may still hear: LLMChain and SequentialChain are legacy (replaced by LCEL), and AgentExecutor is legacy. LangGraph's own create_react_agent still works in 1.x but is deprecated in favour of create_agent.

What a graph adds that a chain cannot express

Cycles — draft, critique, revise; call a tool, look at the result, call another.
Runtime branching — the next step depends on what the model or a tool produced.
Durable state — a checkpointer saves state after every step, keyed by a thread_id, so a run can stop and continue later, even in another process.

When not to use it

If the flow is one prompt and one answer, a graph is only ceremony. If the flow is "model plus tools until done", create_agent gives you the loop, human approval middleware and call limits without writing a graph. Reach for a raw StateGraph when you can name a specific branch, loop or pause that the prebuilt agent does not give you.

A real-life example

An e-commerce marketplace builds a refund assistant in three stages.

Week 1 — it only drafts a polite reply from the order details. One LCEL chain is enough.
Week 3 — it must look up the order, check the return window and call a refund tool. The team uses create_agent with three tools.
Week 6 — finance adds a rule: refunds above Rs 5,000 need a human approver, who may answer hours later, and the refund must never be paid twice if a server restarts. Now the team writes a StateGraph with an approval node that calls interrupt(), and compiles it with a Postgres checkpointer.

Each step up was justified by one new requirement, not by fashion.

Follow-up questions to expect

"Is LangGraph part of LangChain?" — It is a separate package from the same company. LangChain 1.x agents run on the LangGraph runtime, but you can use LangGraph without LangChain.
"What replaced create_react_agent?" — create_agent in langchain.agents. It keeps the same loop and adds middleware for approvals, summarisation, retries and call limits.
"Does LangGraph make the model smarter?" — No. It controls what runs, in what order, and what is saved. The quality of each decision still comes from the model and the prompt.