LangGraph

LangGraph is LangChain Inc.'s low-level orchestration framework and runtime for long-running, stateful agents. Instead of prescribing a prompt format or a single agent architecture, it lets developers represent execution as a graph and mix deterministic application code with model-driven decisions. LangGraph can be used on its own, although LangChain components are commonly used for model and tool integrations. It is aimed at teams that need explicit control over state, routing, retries, side effects and the points where an agent may act.

The core Python repository is MIT-licensed and distributed as the langgraph package. A separate LangGraph.js implementation serves JavaScript and TypeScript users. LangSmith can add tracing, evaluation, visual development and managed deployment, but those services are distinct from the open-source runtime; an application does not need LangSmith or the broader LangChain framework to use LangGraph's orchestration layer.

Top features

• Graph and Functional APIs: Choose explicit nodes, edges and shared state, or add durable execution to ordinary functions with a lower-ceremony API.
• Durable execution: Checkpoints preserve thread state so interrupted or failed work can resume without starting the entire workflow over.
• Two persistence scopes: Checkpointers maintain short-term, thread-level graph state, while stores hold application-defined memory across threads.
• Human-in-the-loop interrupts: Pause at conditional points for approval, editing or input, then resume the same thread from persisted state.
• Streaming and runtime visibility: Stream model messages, state values, node updates, custom events, tasks, checkpoints or debug data.
• Composable control flow: Build branches, loops, parallel paths and subgraphs, and inspect or replay earlier states when troubleshooting.

Use cases

• Gate tool calls, database changes or other consequential actions behind human approval.
• Run research, support or operations agents that may pause, fail and resume over an extended period.
• Coordinate specialist agents through explicit routing, shared state and bounded responsibilities.
• Build custom RAG, SQL or data-processing agents where deterministic validation must surround model decisions.
• Power interactive applications that stream tokens and progress while retaining conversation or workflow state.