LangGraph: building AI agents as graphs

What LangGraph actually does

An AI agent does more than answer a question: it chains steps, calls tools, loops back and sometimes waits for approval. LangGraph models that journey as a graph. Each node is a step (a model call, a tool call, a decision), each edge says where to go next, and a shared state travels between nodes.

That design brings three capabilities the project highlights:

• Durable execution: an agent can run for a long time and resume exactly where it left off after an error.
• Human-in-the-loop: you can pause the agent, inspect its state and correct it before it carries on.
• Memory, both short-term (the current task) and long-term (across sessions).

In practice, teams use LangGraph for business assistants, multi-agent setups where a supervisor hands out work, or processes where the AI must stop and get a green light.

Who it is for, and how to get started

LangGraph is available in Python and in TypeScript (separate repository). It is a low-level orchestration tool: it does not force a ready-made agent architecture on you. That is a plus for experienced developers and a hurdle for newcomers.

A typical first session looks like this: install the langgraph package, follow the quickstart in the official documentation, then define a state, two or three nodes and the transitions between them. LangChain Academy’s free LangGraph course is a useful companion. If you want to move faster, LangChain now points to Deep Agents, a layer on top that provides planning and subagents ready to use.

Strengths and limits

Strengths: precise control over the flow, recovery after failures, human approval checkpoints, no lock-in to one model provider, and support for MCP (a LangGraph agent can be exposed as an MCP tool through LangChain’s Agent Server).

Limits: the learning curve is real, because you have to think in states and transitions. The most comfortable monitoring and deployment path runs through LangSmith, the vendor’s commercial platform. As of October 22, 2026, the LangSmith Plus plan is listed at $39 per seat per month, and LangSmith Deployment is billed on usage. You are not required to use it, but self-hosting then takes more work.

Alternatives

If you would rather write less code, CrewAI thinks in teams of agents with roles. In the Microsoft ecosystem, Microsoft Agent Framework also offers graph-based workflows with checkpointing and human review. For agents built around your documents, LlamaIndex remains a reference. Our agent frameworks comparison helps you place these options side by side.