LangGraph is a low-level orchestration framework for building stateful, controllable AI agents as graphs of steps, with durable execution and human-in-the-loop control.
Created by the LangChain team, LangGraph models agent workflows as graphs where nodes are functions or model calls and edges define control flow. This makes loops, branching, retries and explicit state natural, which is exactly what long-running agents need and what simple chains cannot express.
LangGraph adds production features on top: checkpointing for durable state, streaming of intermediate steps, human approval interrupts, time travel for debugging, and a platform for deploying and monitoring agents at scale. It has become a standard backbone for teams building multi-agent systems that must survive failures and scale to many users. Checkpointers support Postgres, Redis and SQLite, so agent state can live wherever your application data already does. Because state is explicit, teams can pause, inspect and resume runs in ways that prompt-chaining frameworks cannot support.
LangGraph the library is open source and free. The LangGraph Platform for managed deployment, persistence and monitoring is priced per seat and usage, with a free developer tier and enterprise licensing for large-scale or self-managed deployments. The open-source library deploys anywhere, including fully air-gapped environments. The library imposes no fees whatsoever, and self-hosted platforms keep the whole stack under your operational control.
LangGraph is for engineers who need precise control over agent behavior: teams building production agents, multi-agent systems, workflow automation and long-running background workers. If you have outgrown prompt chains or high-level agent builders, this is the layer beneath them. It also suits platform teams building shared agent infrastructure that many product groups will depend upon.
Yes, the core framework is open source and free to use in commercial projects. Paid options only apply to the managed LangGraph Platform and enterprise support.
LangGraph itself is a library, but the LangGraph Platform exposes APIs for launching runs, streaming state and managing checkpoints, so agents can be served as endpoints inside any product.
LangGraph trades ease of use for control: high-level builders get you started faster, but LangGraph gives explicit state, durable execution and debugging tools that those abstractions hide. It is more infrastructure-minded than CrewAI-style frameworks. Choose LangGraph when runs must be recoverable after failure and behavior must be inspectable step by step. LangGraph is infrastructure for serious agents, and the abstraction cost pays for itself as reliability requirements grow.