The production-leaning agent framework when you need state, checkpoints, and human-in-the-loop gates. Python-first with TypeScript support. The graph primitives feel heavy for simple agents but pay off when workflows branch and resume. Hosted "LangGraph Platform" available for managed deployments.
LangGraph is LangChain Inc’s production-leaning agent framework. Where LangChain (the original library) leans toward prompt-chain ergonomics, LangGraph leans toward durable state-machine semantics.
When to reach for it
Multi-step workflows with branching, retries, and human approval gates. The graph model fits naturally with these patterns; you describe the state, the transitions, and the agent fills in the work.
Where it falls short
For simple linear agents, the framework’s weight is hard to justify — you’ll fight the abstractions more than you use them. Python-first heritage shows in the TS SDK’s ergonomics.