LangChain's role in AI agent development

An AI agent is a model working in a loop: it picks a step, calls a tool, reads the result and decides what to do next. Our explainer on what AI agents are covers that loop. LangChain sells the parts around it: code libraries for building the loop, and a paid platform for watching, testing and running it.

The first langchain Python package came out on 24 October 2022. The company says it "started as an open-source side project, without any intention of building a company."[1] Harrison Chase, its chief executive, co-founded the business with Ankush Gola.[2] This profile uses LangChain's own pages for what the products do, and press reports only for the funding.

What the company makes

LangChain's documentation sorts its open-source packages into three layers. Each one is built on the one below it.[3]

• LangGraph is the runtime. It provides durable execution (a run can resume after a crash), persistence, streaming and human-in-the-loop pauses. The documentation calls it "very low-level, and focused entirely on agent orchestration."[4] It can be used without LangChain.[4]
• LangChain is the framework. Its create_agent function runs the standard loop: send a request to the model, run any tool it calls, repeat until it gives a final answer. Middleware hooks let developers change behaviour at each step.[5][6] It runs on LangGraph and offers one interface across model providers such as OpenAI, Anthropic and Google.[6]
• Deep Agents is a harness. It adds ready-made planning, subagents, a virtual filesystem and context compression for long, multi-step tasks.[7]

All three are MIT-licensed on GitHub.[8][9][10] The licence lets anyone use, modify and redistribute the code, including in commercial products.

LangSmith is the commercial product. Its pricing page groups it into observability (traces of each run), evaluation, Engine (which flags issues and proposes fixes), deployment, sandboxes for running agent-written code, an LLM gateway for cost limits and fallbacks, and Fleet, a no-code agent builder.[11] LangSmith's documentation says it "works with many frameworks and providers", including CrewAI and the Vercel AI SDK.[12] Engine v2, Managed Deep Agents and fine-tuning were added in September 2026, so the product list is still changing.[13]

How it fits the agent stack

AI Chronicle

The diagram shows two things that are easy to miss. First, the layers nest. Starting with LangChain or Deep Agents does not shut you out of lower-level control: the 1.0 announcement says developers can "drop down to LangGraph" when they need it.[5] That matters if part of your system should follow a fixed path and part should be left to the model, which is the agent-versus-workflow choice in miniature. LangGraph is designed to mix the two in one graph.[4]

Second, LangChain does not supply the models or the tools. Those come from other vendors and from your own code. LangChain's own comparison page lists rival frameworks (OpenAI Agents SDK, Google ADK, CrewAI, LlamaIndex) and rival runtimes (Temporal, Inngest) in the same categories as its products.[3]

Business model and funding

LangChain runs an open-core model: the libraries are free, and it charges for LangSmith. Fortune described the strategy as keeping early products open source "while creating proprietary platforms."[2] Prices on LangSmith's pricing page, checked 16 October 2026:[11]

On top of that, usage is billed in "LangChain Standard Units" at $1 each. Base traces are kept for 14 days and extended traces for 180.[11] LangChain has not published revenue figures.

Funding, as announced by the company:

• Seed: $10 million led by Benchmark, announced April 2023.[1]
• Series A: $25 million led by Sequoia, announced February 2024, alongside the general release of LangSmith.[14]
• Series B: $125 million at a $1.25 billion valuation, led by IVP, announced 20 October 2025.[15] Fortune and TechCrunch reported the same amount and lead investor, and Fortune names it a Series B.[2][16]

In that announcement, LangChain also said langchain and langgraph have "a combined 90M monthly downloads". It also claimed that 35 percent of the Fortune 500 use its services.[15] These are the company's own figures, and we have not verified them independently.

What to weigh before adopting

Whether you need a framework at all. Fortune notes that OpenAI, Anthropic and Google now offer built-in capabilities that were once LangChain's selling points.[2] For a single model and a few tools, a provider's own SDK may be enough. A framework earns its place when you need to switch models, add middleware or use durable state.

Abstraction weight and change. LangChain has said its old abstractions "were sometimes too heavy." In 1.0 it moved legacy features into a separate langchain-classic package and requires Python 3.10 or later.[5] The promise of "no breaking changes until 2.0" helps, but code written before 1.0 may need migrating.

Where free ends and paid begins. You can run the libraries without paying anything. LangSmith's managed Deployment and Engine start on the paid Plus plan, Fleet's free allowance is small, and self-hosted or hybrid LangSmith requires Enterprise.[11] Price the usage units against your expected trace volume before you commit.

Your data. LangSmith Cloud stores data in the US or the EU, and Enterprise can run in your own cloud network.[11] The pricing page says "LangSmith does not use your data to train models."[11] Check that against your contracts and the terms of service.

Vendor claims are not test results. Product pages describe intended capabilities. They do not show how a tool performs on your tasks, so run your own evaluations. Our guide to why AI agents fail explains what traces help you diagnose.

A worked example: a refund agent

Hypothetical. The company, agent and plan below are invented to show the decision points. Nothing here was built or measured.

A 12-person software company wants an agent that answers billing questions and can issue small refunds.

1. Framework. An engineer uses create_agent with two tools: a read-only invoice lookup and a refund tool. Built-in middleware pauses before each refund call so a human can approve it, and redacts email addresses before text reaches the model.[5]
2. Tracing while building. Tracing goes to LangSmith's free Developer plan (one seat) so the team can see each tool call. The free allowance is 5,000 base traces a month.[11]
3. A fork in the road. Finance wants refunds above a set amount approved over several days. That calls for LangGraph's persistence, so a paused run can resume later without losing its state.[5][4]
4. Before launch. Three people need access, so the plan moves to Plus. The company still has to decide whether its customer data may sit in LangSmith Cloud or needs Enterprise self-hosting.

At each step the team should also price the alternative: a provider SDK plus its own logging. That comparison, not the brand, decides the fit. The same pattern of approval gates and checked sources appears in our tutorial on building a research agent with source verification. If you would rather try the idea without code first, see build your first AI agent without code.

Questions readers commonly ask

Is LangChain free?

The LangChain, LangGraph and Deep Agents libraries are MIT-licensed and free to use.[8][9][10] LangSmith has a free Developer tier, and its paid plans charge per seat plus usage.[11]

What is the difference between LangChain and LangGraph?

LangChain is the higher-level framework with a ready-made agent loop. LangGraph is the lower-level runtime underneath it, used for custom control, long-running work and human approval steps.[3]

Do I have to use LangChain to use LangSmith?

No. LangSmith's documentation says it works with other frameworks and providers.[12]

Is LangChain the same as the company?

No. LangChain Inc. is the company. LangChain is also the name of its framework, and its GitHub repository now describes itself as "The agent engineering platform."[8]

The next step

Before choosing any framework, decide how much of the path the model should control. Read AI agent vs. AI workflow: how to choose.

About these sources

All pages accessed 16 October 2026.

Related reading

• Vapi's approach to voice agents
• Glean and the workplace AI search problem