What is the Graph API in LangGraph, and how do you build a basic graph end-to-end?

What you need to know

1from typing import Annotated, TypedDict2from langchain.chat_models import init_chat_model3from langgraph.graph import StateGraph, START, END4from langgraph.graph.message import add_messages5from langgraph.checkpoint.memory import InMemorySaver67class State(TypedDict):8    messages: Annotated[list, add_messages]910llm = init_chat_model("anthropic:claude-sonnet-4-5")1112def chatbot(state: State) -> dict:13    return {"messages": [llm.invoke(state["messages"])]}1415builder = StateGraph(State)                       # 1. schema16builder.add_node("chatbot", chatbot)              # 2. nodes17builder.add_edge(START, "chatbot")                # 3. edges18builder.add_edge("chatbot", END)19graph = builder.compile(checkpointer=InMemorySaver())   # 4. compile2021config = {"configurable": {"thread_id": "demo-1"}}22graph.invoke({"messages": [{"role": "user", "content": "Hi, I'm Asha"}]}, config)23out = graph.invoke({"messages": [{"role": "user", "content": "What's my name?"}]}, config)24print(out["messages"][-1].content)

It needs pip install langgraph langchain langchain-anthropic and an API key. init_chat_model accepts any "provider:model" string.

What each step does

1. Schema — State has one key, messages, with the add_messages reducer so new messages are appended.
2. Node — chatbot returns only the new message; the reducer adds it to the history.
3. Edges — START to chatbot to END.
4. Compile — checks the wiring and attaches the InMemorySaver checkpointer.
5. Run — the same thread_id on the second call means the graph loads the first turn, so the model knows the name "Asha".

Graph API or Functional API

Both share checkpointers, interrupts and streaming.

A real-life example

A bank's internal team wants a chatbot that answers questions about its leave policy and remembers the conversation. Version one is exactly the graph above plus a retrieve node before chatbot. It took an afternoon. Because it used the Graph API from the start, adding an "escalate to HR" branch a month later was one conditional edge and one new node, not a rewrite of a while loop.

Follow-up questions to expect

• "What does compile() check?" — That there is an edge from START and that every edge points to a node that exists. It does not reject a node that nothing points to, so unreachable nodes need a test or a look at the diagram.
• "What do you get back from compile()?" — A CompiledStateGraph. It supports invoke, stream, batch, ainvoke, astream, get_state and get_state_history.
• "Do you need a checkpointer?" — Only for memory across calls, human-in-the-loop or resume. A one-shot workflow can compile without one.