LangGraph
Agent frameworksA low-level framework for building stateful, multi-step agents as explicit graphs of nodes and edges.
LangGraph models an agent as a graph. Nodes are steps, edges are the decisions between them, and a shared state object travels along as the run proceeds. Loops, branches, retries, and pauses are things you draw rather than things you hope the model handles.
Unlike higher-level "give it a goal and go" frameworks, LangGraph keeps the control flow in your code. That is more to write up front, but when a run misbehaves in production you can point at the node that failed instead of guessing at the model. It is maintained by the LangChain team and works in Python and TypeScript.
Where it's ideally used
Reach for LangGraph when an agent needs real branching — loops, retries, human-in-the-loop pauses — and you want that control flow written down explicitly.
Where it doesn't fit
Overkill for a single prompt-and-response call or a strictly linear chain. The graph ceremony buys you nothing when there are no decisions to make.