# LangChain, from its own documents

slug: what-is-langchain · https://miscsubjects.com/a/what-is-langchain · updated 2026-07-24T16:11:40.212Z

# LangChain, from its own documents

LangChain is the most-starred agent framework on GitHub: 142,500 stars, MIT-licensed, positioned in its own words as "the agent engineering platform." This page is the reference record — what the project says it is, what it actually consists of, and where its boundaries sit — built from its repository and official documentation.

## What it says it is

The repository README puts it in one sentence: "LangChain is a framework for building agents and LLM-powered applications. It helps you chain together interoperable components and third-party integrations to simplify AI application development — all while future-proofing decisions as the underlying technology evolves."

[[embed:source:s1]]

Two commitments live in that sentence. First: LangChain is a *framework* — a library you import into your own Python or TypeScript process, not a service you call. Nothing runs until you host it. Second: its core value claim is interchangeability — swap models, vector stores, and tool providers without rewriting the application around them.

## The harness, not the model

The current documentation reduces the whole product to one abstraction: "LangChain provides create_agent: a minimal, highly configurable agent harness." The docs' own formula is "Agent = Model + Harness" — the model does the reasoning, and the harness is everything wrapped around the loop: prompts, tools, and behavioral middleware.

[[embed:source:s2]]

That is a real position in a live design argument. LangChain is betting that the durable engineering surface is the loop around the model, not the model. Their product tiering says the same thing three ways: Deep Agents for batteries-included use, LangChain for customizable harnesses, LangGraph for "advanced deterministic and agentic workflow combinations."

## LangGraph is the load-bearing layer

Since the 1.0 reorganization, LangChain agents are "built on top of LangGraph," the project's "low-level orchestration framework for building stateful agents" — 38,100 stars on its own. LangGraph models an agent as a graph: nodes are steps, edges are transitions, and the runtime adds the three things production teams actually pay for — durable execution that resumes after failure, human-in-the-loop inspection of agent state mid-run, and both short-term and long-term memory. Klarna, Replit, and Elastic run on it.

[[embed:source:s3]]

Read the dependency direction closely and the famous name inverts: "LangChain" is now the friendly wrapper; LangGraph is the engine. Someone evaluating the ecosystem in 2026 is really evaluating LangGraph.

## What it is not

The boundary matters as much as the feature list. LangChain is not a hosted service: you deploy, scale, and secure the process it runs in (LangSmith, the company's commercial observability product "for developing, debugging, and deploying AI agents," is where hosting concerns move — and where the open-source project meets the paid one). It is not a tool registry: integrations are Python packages you install, not rows you discover at runtime. It is not a protocol: two LangChain applications have no standard way to talk to each other — that gap is what MCP and A2A exist to fill. And it holds no opinion about authorization: what an agent may do is whatever the process's credentials can do.

Those four absences are not criticisms. They are the shape of the category "framework" — and the reason evaluating any specific system against LangChain requires saying which of the four gaps that system fills.


## Sources

1. langchain-ai/langchain (GitHub) — https://github.com/langchain-ai/langchain
2. LangChain official docs — overview — https://docs.langchain.com/oss/python/langchain/overview
3. langchain-ai/langgraph (GitHub) — https://github.com/langchain-ai/langgraph