What it is
LangGraph is LangChain's low-level framework for building agents as graphs: nodes, edges and a shared state saved at every step. That enables durable execution, memory, streaming and pauses so a person can approve or correct the agent. It's open source (MIT), available in Python and JavaScript, and deploys anywhere you like or on LangSmith.
How to use it
- Model your agent as a graph, with a node per step and conditional branches.
- Persist state with checkpoints to resume long runs after a failure.
- Pause for human review before sensitive actions, then continue.
- Stream in real time the agent's tokens and steps to your UI.
- Build multi-agent systems, hierarchical or networked, with the same framework.
What it's good for
- A fintech using an agent to review loan applications, with human approval before replying.
- An engineering team that needs long-running agents able to survive crashes and restarts.
- A software company coordinating several specialist agents under a supervisor.
Pros
- Fine-grained control: you define every node, branch and state
- Durable execution with checkpoints and resumption
- Built-in human-in-the-loop review
- Open source (MIT) in Python and JavaScript
Cons
- Steep learning curve: you have to think in graphs and state
- More code than a high-level framework for simple agents
- Managed deployment and observability are paid separately through LangSmith
