What it is
LangGraph lets you build an agent as a graph: nodes are steps (a model call, a tool, your own function) and edges decide what runs next, including loops and branches. State is saved at each step by a checkpointer.
That saved state gives it durable execution (a run can stop and resume after a crash), memory across sessions, streaming, and human-in-the-loop pauses where a person approves or edits before the agent continues. Version 1.0 shipped in October 2025 with no breaking changes, and the 1.2 line was current in October 2026. A JavaScript port, LangGraph.js, is maintained separately.
It is made by LangChain but can be used on its own. LangChain's create_agent and the higher-level Deep Agents package are built on it. Tracing, hosting and the Studio debugger are sold separately through LangSmith.
Our review
LangGraph is the framework to reach for when an agent's control flow has to be inspectable and resumable, and overkill for anything simpler. It has about 43,800 GitHub stars, released 1.2.16 on 16 October 2026, and in 2026 comparisons it remains the reference point that other orchestration frameworks are measured against.
Its strength is control. You decide which step runs next, what state is kept and where a person must sign off. Durability depends on the checkpointer you choose: the docs warn that InMemorySaver loses everything on restart and point production users to PostgresSaver, with a reminder to prune old checkpoints and keep thread IDs under 255 characters. The 1.0 release froze the core graph API, which matters to teams already running it.
The cost is effort. Defining state schemas, nodes and edges is more code than a plain agent loop, and newcomers report a steep learning curve. The README now steers people who want results quickly to Deep Agents, a higher-level package built on LangGraph, and for a basic tool-calling agent LangChain's create_agent or the OpenAI Agents SDK gets there faster. TypeScript teams often find Mastra more natural than LangGraph.js.
The commercial pull is towards LangSmith. Plus costs $39 per seat a month, includes one small serverless deployment, then bills runtime at 0.0675 LSU (one LSU is $1) per vCPU-hour; self-hosting LangChain's deployment platform is Enterprise only. Most of the 13 security advisories published against the repo hit the SDK's custom-auth and store layers rather than the graph engine, including a high-severity flaw in August 2026 where resource decorators silently ignored their actions= restriction.
The library is free; the parts that make it comfortable in production, Studio, tracing and managed persistence, are where LangChain earns its money.
Where it earns its place
A fintech team building a loan-document review agent
Have an agent draft a decision but pause for a credit officer's sign-off
An interrupt stops the graph before the decision node, state is stored in Postgres, and the officer edits or approves before the run resumes. The library is free; the cost is Postgres and the engineering time to model the state schema.
A platform engineer running multi-hour research agents on Kubernetes
Keep long agent runs alive through pod restarts and deploys
Durable execution resumes from the last checkpoint after a crash. Self-managed, that means running your own Postgres; on LangSmith Deployment it is $39 per seat plus compute billed per vCPU-hour, and self-hosting that platform requires an Enterprise contract.
A product manager choosing a framework for a support-automation pilot
Ship a first agent quickly without boxing the team in later
Start with create_agent or Deep Agents, which run on LangGraph underneath, and drop to raw graphs only for the steps that need explicit branching or approvals. That keeps the pilot cheap while leaving a path to the lower-level API.
Before you commit
- Setup
- pip install -U langgraph for Python, or the separate LangGraph.js package; the quickstart and LangChain Academy's free course cover a first graph. Production persistence needs a checkpointer database such as Postgres.
- Integrations
- Works with any model or tool through plain Python functions, and with LangChain's integrations when installed; LangSmith adds tracing, evaluation, Studio and managed deployment.
- Data and privacy
- MIT licence with no commercial-use limits. State lives in whichever checkpointer you configure; traces go to LangSmith only if you set it up, and LangSmith Developer and Plus data sits in LangChain's US or EU cloud. 13 GitHub advisories so far, five rated high.
- Cost at scale
- Self-managed, the cost is your compute and Postgres. On LangSmith, Plus is $39 per seat a month with 10,000 base traces, then pay-as-you-go traces and deployment compute at 0.0675 LSU per vCPU-hour runtime and 0.177 LSU per vCPU-hour for the database.
- Limits
- 511 open issues and pull requests. Checkpoints accumulate unless pruned, and some features in the README (Studio, managed deployment) are LangSmith products rather than part of the MIT library.
What's good
- Fine-grained control over agent flow, loops and branches
- Checkpointing with Postgres for durable, resumable runs
- Human-in-the-loop interrupts are first-class
- Core graph API stable since 1.0
What's not
- More code and concepts than simple agent SDKs
- Steep learning curve; the README itself points beginners to Deep Agents
- Managed deployment and the best debugging tools are paid LangSmith features
Best for
Engineering teams building long-running or multi-step agents that must be reliable and auditable.
Alternatives
LangGraph: common questions
- Is LangGraph worth it?
- Explicit control over every step an agent takes, paid for in code and learning time, and worth it once an agent must survive restarts and wait for a human. We rate it 4.6 out of 5.
- How much does LangGraph cost?
- Free and open source under the MIT licence. LangSmith (tracing and deployment) is separate: Developer free for one seat, Plus $39/seat/month plus compute, Enterprise custom.
- What are the downsides of LangGraph?
- More code and concepts than simple agent SDKs; steep learning curve; the README itself points beginners to Deep Agents; managed deployment and the best debugging tools are paid LangSmith features.
- Who is LangGraph best for?
- Engineering teams building long-running or multi-step agents that must be reliable and auditable.
- How much does LangGraph cost for a team or in production?
- Self-managed, the cost is your compute and Postgres. On LangSmith, Plus is $39 per seat a month with 10,000 base traces, then pay-as-you-go traces and deployment compute at 0.0675 LSU per vCPU-hour runtime and 0.177 LSU per vCPU-hour for the database.
- Is LangGraph safe for company data?
- MIT licence with no commercial-use limits. State lives in whichever checkpointer you configure; traces go to LangSmith only if you set it up, and LangSmith Developer and Plus data sits in LangChain's US or EU cloud. 13 GitHub advisories so far, five rated high.
- What are the best LangGraph alternatives?
- The closest alternatives we have reviewed are LangChain, CrewAI, Microsoft Agent Framework and Google Agent Development Kit (ADK).
Answers checked Oct 23, 2026.
Make LangGraph? Show this rating on your site
