Overview
LangGraph is a low-level orchestration framework and runtime for long-running, stateful AI agents. It gives developers fine-grained control over agent behavior by representing workflows as graphs and combining predictable application logic with model-driven decisions.
The framework focuses on orchestration rather than abstracting prompts or prescribing an agent architecture. LangChain components can supply models and tools, but LangGraph does not require LangChain and can serve as the underlying runtime for bespoke agent systems.
Key Capabilities
- Mix deterministic, hand-coded steps with flexible LLM-driven steps in one graph.
- Support durable execution for workflows that need to continue reliably over time.
- Stream intermediate output and agent activity.
- Preserve state and memory across a workflow.
- Add human-in-the-loop checkpoints where oversight is required.
This combination lets teams keep critical logic predictable and auditable while applying AI only where adaptable reasoning is useful.
Position in the LangChain Stack
LangGraph provides the orchestration layer beneath higher-level agent frameworks and harnesses. LangChain offers configurable abstractions and integrations for models, tools, prompts, middleware, and common agent loops, while Deep Agents adds planning, subagents, filesystem tools, and context management on top of LangGraph.
LangSmith complements these tools with tracing, evaluation, prompts, and deployment capabilities. Its tracing features can help developers debug agent behavior and evaluate outputs, while LangSmith Engine monitors traces, detects issues, and proposes fixes.
Getting Started
The JavaScript packages are @langchain/langgraph and @langchain/core. A basic graph defines a state schema, adds nodes that read or update that state, connects the start and end points with edges, compiles the graph, and invokes it with an initial state.
Because LangGraph is deliberately low-level, familiarity with models and tools is recommended before adoption. Developers who want prebuilt architectures for standard LLM and tool-calling loops can instead begin with LangChain agents, while LangGraph is suited to applications that need direct control over stateful orchestration.