How to Build Stateful Agent Workflows with LangGraph

What you need

Python 3.10+ installed.

LangGraph installed: `pip install langgraph`.

LangChain installed for the tool ecosystem: `pip install langchain`.

An API key for your model provider.

Step 1: Define your agent state

LangGraph uses a typed state object that flows through every node in the graph.

Define the state schema: messages (conversation history), current_step, research_findings, draft, and approval_status.

The state is the single source of truth — every node reads from and writes to it.

Step 2: Create nodes for each step

Each node is a function or callable that takes the state, performs work, and returns updates.

Nodes: `research_node` (calls web search), `draft_node` (writes content), `review_node` (checks quality), `publish_node` (finalizes).

Nodes can have different models, tools, and prompts.

Step 3: Add edges and routing logic

Edges define the flow between nodes. Simple edges go from one node to the next.

Conditional edges use routing functions: 'If review approves, go to publish_node. If not, go to draft_node for revision.'

This is where LangGraph's graph model shines compared to linear chains.

Step 4: Enable persistence for long-running workflows

Add a checkpointer (SQLite or Postgres) so the workflow state persists across interruptions.

This allows workflows to run for hours or days, surviving server restarts.

Persistence is essential for human-in-the-loop workflows where a human may take hours to respond.

Step 5: Configure human-in-the-loop checkpoints

Insert interrupt nodes where the workflow pauses and waits for human input.

The human reviews the current state, provides feedback, and the workflow resumes from that point.

LangGraph supports approval, rejection, and edit actions at checkpoints.

Step 6: Deploy the workflow

LangGraph workflows can be deployed as LangGraph Server, as a standalone API with FastAPI, or embedded in your application.

Add monitoring and logging to track workflow execution, state changes, and error rates.

Scale horizontally by running multiple workflow instances across worker processes.

FAQ

Build LangGraph Workflows

1. Install LangGraph: pip install langgraph
2. Define a typed state schema
3. Create function nodes for each step
4. Add edges with conditional routing
5. Enable persistence with SQLite/Postgres
6. Add human-in-the-loop checkpoints