LangGraph
★ 36.3k · GitHub →Low-level orchestration framework for stateful, long-running agents modeled as graphs.
Install
$ pip install langgraph What is LangGraph?
LangGraph is the LangChain team's answer to a hard question: what happens when your agent needs to run for hours, survive crashes, and pause for human approval? You model the workflow as a graph — nodes are steps, edges are transitions, and a checkpointer persists the full state after every node. That checkpoint architecture is what enables its signature features: durable execution that resumes exactly where it left off, time-travel debugging, and human-in-the-loop interrupts that can hold state for days.
It is deliberately low-level. Unlike CrewAI or AutoGen, it does not hand you prefab agent roles or conversation patterns; you wire the state schema, nodes, and conditional edges yourself. That is more upfront work, but it means the control flow is exactly what you wrote — no framework magic deciding when an agent speaks. Companies like Klarna, Replit, and Elastic run production agents on it, and LangChain's own higher-level agent abstractions compile down to LangGraph under the hood.
The cost is a real learning curve: thinking in state reducers and graph compilation is unfamiliar if you have only written sequential agent loops, and simple tasks feel over-engineered. LangGraph itself is MIT open source and works standalone (LangChain is optional); LangSmith and the managed deployment platform are the commercial layer on top.
Key capabilities
- ▸ Graph-based control flow: explicit nodes, conditional edges, cycles, and parallel branches
- ▸ Checkpointer-backed durable execution — resume from exact state after crashes or restarts
- ▸ Human-in-the-loop interrupts that pause a run mid-graph and hold state indefinitely
- ▸ Built-in short-term and long-term memory primitives for stateful agents
- ▸ Standalone MIT core with optional LangSmith tracing and managed deployment
When to use it
Reach for LangGraph when reliability is the requirement: multi-step workflows that must survive restarts, approval gates in the middle of a run, or agent state you need to inspect and replay. It is the strongest choice in this list for production durability. Do not pick it for a weekend prototype or a single tool-calling agent — the graph ceremony will slow you down, and a simple loop in Pydantic AI or the OpenAI Agents SDK ships the same thing in a tenth of the code.
Alternatives
Frequently asked questions
Do I need LangChain to use LangGraph?
No. LangGraph is a standalone MIT-licensed library — you can call any model SDK directly inside your graph nodes. Many teams use LangGraph for orchestration while skipping LangChain entirely. That said, LangChain's model and tool integrations plug in conveniently if you want them, and the two are designed to compose.
What is a checkpointer in LangGraph?
A checkpointer persists the graph's full state after each node executes, to a backend like SQLite, Postgres, or Redis. This is what enables durable execution (resume after a crash), human-in-the-loop interrupts (pause and wait for approval), and time-travel debugging (rewind to any prior state and re-run from there).
Is LangGraph overkill for simple agents?
Usually, yes. If your agent is a single loop of "call model, run tools, repeat," defining state schemas and compiling graphs adds ceremony without benefit. LangGraph earns its complexity when you need branching workflows, persistence across restarts, or pausing for human input. Start simpler and migrate when those needs appear.
Star count updated 2026-07-02 · refreshed weekly from the GitHub API · browse all tools →