langchain-ai/langgraph

What LangGraph does

LangGraph is a low-level orchestration framework for building stateful, long-running agents and workflows. Instead of hiding control flow behind a high-level abstraction, it exposes the graph-level primitives needed to coordinate steps, branch, persist state and resume execution after failures. The core capabilities are durable execution, human-in-the-loop interrupts, short- and long-term memory, LangSmith-based tracing, and deployment options for long-running workflows (why use LangGraph).

It is aimed at developers who need explicit control over agent behavior and state, rather than a ready-made agent product. The Python package is the primary distribution; a JS/TS equivalent is maintained separately (LangGraph.js). The project is built by LangChain Inc. but can be used without the LangChain framework (README).

Install and start

LangGraph requires Python and is installed from PyPI:

pip install -U langgraph

After installation, the official quickstart is the shortest path to a working graph: it walks through defining state, adding nodes and edges, and running the graph (LangGraph Quickstart). Concept overviews and guides are in the main documentation (docs), and API-level details are in the Python reference.

The README does not include a complete runnable example, so the install command alone does not produce a working agent; follow the quickstart for the graph construction steps. For a higher-level package with planning and subagents, the project points to Deep Agents.

Constraints and related components

LangGraph is MIT-licensed (license). It is a framework, not a hosted service: durable execution, memory and human-in-the-loop behavior are features you wire into your own application; production deployment and observability are handled through LangSmith (LangSmith, deployments). Debugging and tracing requirements therefore depend on whether you use LangSmith or another integration.

For teams already using LangChain, the components integrate directly; for others, the optional LangChain relationship should not be assumed to be required (ecosystem). The JS/TS version has its own package and documentation, so Python-specific examples do not carry over directly.

HotRepo's take

The only recorded HotRepo snapshot for this repository is 2026-11-13 on the daily Python list, showing 44,895 stars, 7,743 forks and 56 stars gained that day, with a best position of 13. That single record establishes a position on the Python daily list on that date, not a broader trend. The technical facts that stand out are the low-level positioning and the explicit separation between the MIT-licensed framework and the paid LangSmith observability and deployment products: the repository provides the orchestration layer, while production tracing and hosting are separate concerns. Readers should evaluate it as infrastructure for stateful workflows, not as a complete agent product.