LangGraph for production agents

Artificial intelligence agents are software systems that call models and tools in steps, and LangGraph packages those steps as a graph that can pause, branch, loop, and resume in production. (docs.langchain.com) LangChain’s documentation says LangGraph is a low-level orchestration framework for “long-running, stateful agents,” with durable execution, streaming, and human review built in. Its GitHub repository describes the project as infrastructure for building, managing, and deploying stateful agents. (docs.langchain.com) (github.com) The basic model is simple: state is the shared memory, nodes do work, and edges decide what runs next. LangChain’s graph application programming interface docs say those edges can be fixed transitions or conditional branches, which lets a workflow route itself based on the current state. (docs.langchain.com) That structure is aimed at a production problem: many agent systems do not run in one straight line. LangChain’s JavaScript guide draws a line between workflows, which follow predetermined code paths, and agents, which choose tools and steps dynamically during execution. (docs.langchain.com) LangGraph’s own product page says developers can build single-agent, multi-agent, and hierarchical control flows in one framework. The same page says it supports human approval checks and token-by-token streaming so users can see actions unfold in real time. (langchain.com) The company also positions LangGraph below its higher-level agent tools. The Python overview says developers who want a faster start can use LangChain agents, while LangGraph is the lower-level layer focused on orchestration rather than prompts or a fixed architecture. (docs.langchain.com) That split shows up in the codebase. The main Python repository had about 28,900 GitHub stars and the JavaScript repository had about 2,800 stars as of April 12, 2026, and both repositories describe LangGraph as the control layer for more customizable agent systems. (github.com 1) (github.com 2) The production pitch is no longer limited to LangChain’s own materials. Amazon Web Services published a January 13, 2026 post describing a DynamoDB-backed checkpoint system for LangGraph, aimed at storing agent state so workflows can recover from failures and keep running across long sessions. (aws.amazon.com) LangChain’s site says companies including Klarna, Uber, J.P. Morgan, Replit, LinkedIn, and GitLab use LangGraph, though those references come from the company’s own marketing pages rather than independent audits. The through line is consistent: the selling point is not a smarter model, but tighter control over how model calls, tools, memory, and human approvals fit together after deployment. (docs.langchain.com) (github.com) The reason LangGraph keeps surfacing in agent discussions is practical. If an agent is going to run for minutes, call several tools, hand work to another agent, or wait for a person to approve an action, it needs a state machine more than a chat box, and that is the layer LangGraph is built to provide. (docs.langchain.com) (langchain.com)