Last verified
LangGraph
VerifiedBuild stateful, cyclical agent workflows as explicit graphs with LangChain
Agent details
- Kind
- framework
- Framework
- LangGraph
- Autonomy
- supervised
Tools
- tool_node
- human_in_the_loop
- state_checkpointing
"Cycles and checkpoints. LangGraph handles the patterns that ReAct loops cannot."
What is LangGraph?
LangGraph is a LangChain extension library for building stateful agent workflows as explicit directed graphs. Each node is an agent step (LLM call, tool use, conditional branch) and edges define execution flow. The graph model supports cycles and persistent state, making it a developer tools framework for complex agentic patterns that require loops, self-correction, and multi-turn memory.
How does it work?
You define a StateGraph with a typed state schema, add nodes as Python functions that read and update the state, and connect them with edges, including conditional edges that branch based on state values. LangGraph compiles the graph into a runnable and executes it, managing state checkpointing at each step. Human-in-the-loop breakpoints can pause execution at designated nodes for operator confirmation before continuing.
For teams who prefer a role-based, higher-level abstraction over explicit graph definitions, CrewAI provides a crew/agent/task model that handles common multi-agent patterns with less graph-definition code. LangGraph gives more control when those patterns do not fit.
When should you use it?
Use LangGraph when you need fine-grained control over agent execution flow: conditional branching, retry loops, reflection steps, or multi-agent handoffs that require explicit state transitions. It suits production agent systems where you need to inspect, debug, and resume runs reliably.
Frequently asked questions
What is the key advantage of LangGraph's graph model over a linear chain?
LangGraph supports cycles: an agent can loop back to an earlier node based on a condition, enabling retry logic, tool-use loops, and reflection patterns where the agent checks its own output before proceeding. Linear chains execute in one direction only and cannot model agent self-correction.
How does LangGraph handle persistent state between graph steps?
LangGraph uses a typed State object passed between every node. Nodes read from and write to this shared state, and LangGraph can checkpoint the state to a database (SQLite, PostgreSQL, Redis) so long-running agents can be paused and resumed without re-running earlier steps.
Can LangGraph be used with Claude models specifically?
Yes. LangGraph is LLM-agnostic and integrates with any LangChain-compatible model, including Anthropic Claude via `ChatAnthropic`. Use `bind_tools` on the Claude model instance to enable tool calling within graph nodes. Claude's extended context window is useful for graphs that accumulate large state objects.