LangGraph

LangGraph is a low-level orchestration framework and runtime for building, managing, and deploying long-running, stateful agents. It is especially useful when an agent workflow needs durable execution, state, memory, checkpoints, human-in-the-loop, or multi-agent orchestration.

Primary reference: https://docs.langchain.com/oss/python/langgraph/overview

Where LangGraph fits

flowchart TB
    A[LLM / Model] --> B[LangGraph]
    B --> C[State graph]
    C --> D[Nodes]
    C --> E[Edges]
    C --> F[Checkpointing]
    C --> G[Human-in-the-loop]
    C --> H[Agent service]

LangGraph answers:

How do I orchestrate stateful, long-running, controllable agent behavior?

It does not primarily answer:

How should my software team approve specs, track delivery, or run code review?

Core concepts

Concept	Role
State	Shared data that moves through the graph
Node	A unit of work, often a model/tool/function step
Edge	Routing from one node to another
Checkpoint	Durable saved execution state
Human-in-the-loop	Human review or input during execution
Memory	Persistent context across interactions
Multi-agent graph	Multiple specialized agents coordinated by a graph

Basic state graph

stateDiagram-v2
    [*] --> Plan
    Plan --> ToolUse
    ToolUse --> Reflect
    Reflect --> Plan: Need more work
    Reflect --> Final: Done
    Final --> [*]

Multi-agent orchestration

flowchart TB
    U[User task] --> S[Supervisor node]
    S --> R[Research agent]
    S --> C[Coder agent]
    S --> T[Test agent]
    R --> S
    C --> S
    T --> S
    S --> F[Final result]

When to use LangGraph

Use LangGraph when:

• A simple chain is not enough.
• You need stateful multi-step workflows.
• Agent execution may be long-running.
• You need human-in-the-loop review.
• You need durable checkpoints.
• You need multi-agent coordination.
• You need runtime control and observability for an agent app.

When not to use LangGraph

Avoid LangGraph when:

• a simple prompt/tool call is enough;
• a basic LangChain agent is enough;
• your problem is software delivery process, not app runtime orchestration;
• your team does not need state, checkpoints, or graph control.

LangGraph vs LangChain

Dimension	LangChain	LangGraph
Best for	Simple-to-medium AI apps and agents	Stateful, long-running, controllable agents
Abstraction	Higher-level app/agent building blocks	Lower-level graph orchestration
State	Available but not the main mental model	Central concept
Human-in-loop	Possible	First-class pattern
Durable execution	Not the main focus	Core strength
Multi-agent	Possible	Stronger fit

Step-by-step: support agent

1. Define support workflow states:
   • classify request;
   • retrieve customer context;
   • propose answer;
   • escalate if needed;
   • human approval for risky replies.
2. Model state schema.
3. Create nodes.
4. Add edges and routing logic.
5. Add checkpointing.
6. Add human-in-the-loop nodes.
7. Add tool permissions.
8. Add tests/evals per node and full graph.
9. Add monitoring and tracing.
10. Deploy as an agent service.

Definition of done

A LangGraph agent is production-ready when:

1. State schema is explicit.
2. Node responsibilities are clear.
3. Edges and fallback paths are tested.
4. Checkpoints are configured.
5. Human-in-the-loop gates are defined.
6. Tool permissions are safe.
7. Observability exists.
8. Evals cover important scenarios.