LangGraph: Stateful Multi-Step Agent Graphs

After this lesson, you will be able to: Use LangGraph to build graph-based agentic workflows with conditional branching, cyclic loops, and persistent checkpointing.

AgentExecutor is fine for simple loops. Real workflows have conditionals, branches, parallel steps, and human-in-the-loop checkpoints. LangGraph models agents as state graphs, nodes are functions, edges are control flow, and state flows between them. This lesson builds a multi-step research agent with branching.

Prerequisites:LangChain: The Agent Framework

Why graph?

AgentExecutor = loop. LangGraph = graph. Same model deciding what to do, but you control where it can go next, what state persists, when to stop. Production agent code lives in LangGraph today.

START

→

plan

→

research

→

reflect

→

write

→

END

A LangGraph state machine. research and reflect loop until the plan is satisfied (research can even call itself), then a conditional edge sends control to write.

Install + concepts

1
1. `pip install langgraph langchain-anthropic`
2
2. Mental model: a graph has State (the dict that flows through), Nodes (functions that read+update state), and Edges (where to go next, conditionally).
3
3. You compile the graph; calling it runs from START until END.

Plan-research-write graph

Three nodes, conditional edges, persistent state.

python

from typing import TypedDict, Annotated
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from langchain_anthropic import ChatAnthropic

model = ChatAnthropic(model='claude-sonnet-4-6')

class State(TypedDict):
    topic: str
    plan: str
    research: list[str]
    draft: str

def plan(state: State):
    p = model.invoke(f'Make a 3-bullet research plan for: {state["topic"]}').content
    return {'plan': p}

def research(state: State):
    r = model.invoke(f'Plan: {state["plan"]}\nResearch this and return facts.').content
    return {'research': state.get('research', []) + [r]}

def write(state: State):
    d = model.invoke(f'Plan: {state["plan"]}\nFacts: {state["research"]}\nWrite a 300-word piece.').content
    return {'draft': d}

g = StateGraph(State)
g.add_node('plan', plan)
g.add_node('research', research)
g.add_node('write', write)
g.add_edge(START, 'plan')
g.add_edge('plan', 'research')
g.add_edge('research', 'write')
g.add_edge('write', END)
app = g.compile()

result = app.invoke({'topic': 'why ravens are smart'})
print(result['draft'])

Checkpointing for resumability

Pass `checkpointer=MemorySaver()` (or PostgresSaver for prod) to `g.compile()`. Each step's state is saved. You can resume mid-flow, replay, or fork, invaluable for long-running agents.

←LangChain: The Agent Framework

Back to Code-First Agents

LlamaIndex: RAG Pipelines and Data Agents→