LangChain: the original framework for building AI apps

From early pioneer to agent toolkit

LangChain first shipped in October 2022, a few weeks before ChatGPT launched, and quickly became one of the first libraries for programming around a large language model. The original idea was to “chain” components (a model, a prompt, a data source, a tool) instead of wiring everything by hand.

The project has changed a lot since then. Today the company describes it as a framework for building agents and AI applications, with a standard interface for models, embeddings and vector stores. One key point: according to the official docs, LangChain agents now run on LangGraph, which gives them failure recovery, persistence and human approval steps.

What you can build with it

• An agent that decides which tools to call (search, calculations, an internal API) to finish a task.
• An assistant that answers from your own documents using RAG, retrieval-augmented generation.
• A chatbot that can move from one model provider to another without a heavy rewrite.

LangChain is available in Python and in JavaScript/TypeScript (separate repository). Its catalog of integrations with models, tools and databases is one of its biggest assets.

Getting started

The official docs suggest a three-step start: install the langchain package with your usual Python package manager, initialize a model by naming its provider and model ID, then create an agent and hand it a few tools. You will need an API key from the model provider. If you want a more autonomous agent from day one, the company now points to Deep Agents, a layer on top that adds planning and subagents.

Strengths and weaknesses

Strengths: a very broad integration ecosystem, freedom to pick your model, plenty of documentation, a natural path to LangGraph as your project grows, and LLM observability through LangSmith.

Weaknesses: the library has been reworked many times, so older tutorials often no longer match the current API; always check the date of what you are reading. The abstraction layers can also hide what is really going on, which makes debugging harder without a tracing tool. Finally, the most integrated tooling (LangSmith) is a commercial product.

Alternatives

For agents focused on working with documents, LlamaIndex is the most direct alternative. For teams of agents organized by role, take a look at CrewAI.