LangGraph

Low-level graph-based agent orchestration from the LangChain team — explicit state, loops, and human-in-the-loop control.

framework open-source Updated 2026-09-23

Use when

  • Engineering teams who need fine control over agent control flow and state
  • Long-running agents with loops, branching, and human-in-the-loop steps
  • Teams already in the LangChain ecosystem wanting low-level orchestration primitives

Operational upside

  • Explicit control over agent flow — nodes, edges, and shared state instead of a hidden agent loop
  • Graph state model handles cycles, branching, and human-in-the-loop pauses cleanly
  • Checkpointing and persistence through a configured checkpointer for resumable agents
  • Model-agnostic — works with OpenAI, Anthropic, open, and local models via adapters
  • MIT licensed and self-hostable; runs as a Python or TypeScript library

Avoid or plan for

  • Code-first with a real learning curve — non-developers should not start here
  • Lower-level than most frameworks; you assemble more of the plumbing yourself
  • Naming overlap with LangChain confuses newcomers about which to use when
  • Smaller ready-made connector catalog than the LangChain framework it sits under
  • Overkill for a single LLM call or a simple linear chain

What it is

LangGraph is a low-level orchestration framework from the LangChain team for building stateful, controllable agents. It models an agent as a graph of nodes and edges over shared state. With a configured checkpointer, it can persist checkpoints for loops, branching, human-in-the-loop pauses, and recovery — the control-flow primitives higher-level frameworks tend to hide.

Who it’s for

Choose LangGraph when you think in graphs and want to own every step of the agent’s control flow: backend and platform teams building long-running, resumable agents, and teams already on LangChain that have outgrown the default agent loop. If you want a visual canvas or a one-line agent, start elsewhere.

Frontier verification — checkpointing is not exactly-once execution

LangGraph’s persistence documentation describes checkpoints associated with execution threads. That can support resume, inspection, and human-in-the-loop flows. It does not make a payment, database write, or message delivery exactly once.

Temporal now documents an integration path for LangGraph workloads. Add that runtime only when durable coordination across failures is a separate requirement from graph-state persistence. Activities may retry, so external writes still need idempotency.

Decision rule: use LangGraph checkpoints for graph progress; add durable execution only when the failure and operational boundary justify a second runtime.

Current sources: LangGraph persistence and Temporal’s LangGraph integration brief. Verified 2026-09-13.

Security note — verified 2026-09-23

An official High advisory affects the Python langgraph-sdk package, versions 0.1.45 through 0.4.3: actions= restrictions on certain resource-scoped custom-auth decorators can be applied incorrectly. Only deployments using those affected Python authorization patterns are in scope; this is not a claim about every LangGraph application or the framework’s version. The fix is in langgraph-sdk 0.4.4. The advisory lists no workaround, so affected deployments should upgrade.

Source: LangGraph SDK custom-auth security advisory.

FAQ

Is LangGraph free?
LangGraph has a free tier or open-source edition. See pricing details on the official site for paid features and usage limits.
What is LangGraph best for?
Engineering teams who need fine control over agent control flow and state Long-running agents with loops, branching, and human-in-the-loop steps Teams already in the LangChain ecosystem wanting low-level orchestration primitives
What are the main downsides of LangGraph?
Code-first with a real learning curve — non-developers should not start here Lower-level than most frameworks; you assemble more of the plumbing yourself Naming overlap with LangChain confuses newcomers about which to use when
Who should use LangGraph?
Low-level graph-based agent orchestration from the LangChain team — explicit state, loops, and human-in-the-loop control. See our review for the full pros and cons.