Overview
LangGraph is a low-level orchestration framework and runtime for stateful, long-running agents. It is intended for developers who need fine-grained control over agent behavior rather than a prebuilt, high-level agent architecture.
A LangGraph workflow can combine deterministic, hand-coded logic with flexible LLM-driven steps in the same graph. Although its documentation commonly uses LangChain components for models and tools, LangGraph does not require LangChain.
Key Capabilities
- Durable execution supports long-running workflows that preserve state.
- Streaming exposes intermediate activity while an agent runs.
- Persistence and memory retain information needed by stateful agents.
- Human-in-the-loop support allows people to participate in controlled workflows.
- Graph orchestration combines predictable steps with model-driven decisions.
- Low-level primitives leave prompts and architecture under developer control.
How It Fits the Stack
LangGraph serves advanced orchestration needs where an application must precisely control how deterministic and agentic work interact. LangChain offers higher-level abstractions and integrations for models, tools, prompts, middleware, and common agent loops, while LangGraph supplies the underlying orchestration runtime used by LangChain agents.
Managed Deep Agents provides a hosted harness with planning, subagents, a virtual filesystem, and long-term memory. The Deep Agents SDK offers the same style of harness with self-hosting flexibility, while LangSmith focuses on tracing, debugging, evaluation, deployment, and monitoring.
Getting Started
The Python package can be installed with pip or uv. A basic workflow defines state, adds a node that updates it, connects that node between start and end points, compiles the graph, and invokes it with an initial message.
Developers new to agents are advised to first understand models and tools. Those who want a higher-level starting point can use LangChain agents, while LangGraph is suited to bespoke workflows that require explicit control, auditability, persistence, streaming, or human participation.
