LangChain

1 · What it is

LangChain is an open-source framework for building apps and agents that run on large language models (LLMs). A framework is a set of ready-made building blocks you plug your own code into. LangChain is released under the MIT license. There is also a version for JavaScript and TypeScript, called LangChain.js.

Imagine you want an app that answers questions using an AI model. A provider is a company that hosts AI models and offers them through an API, such as OpenAI, Anthropic or Google. LangChain puts one standard interface in front of many providers. An interface is the set of commands your code uses. Because it stays the same, switching models needs only small code changes. Many providers have their own LangChain package, such as langchain-openai or langchain-anthropic. The docs list more than 1,000 integrations. These are ready-made connections to chat and embedding models, tools, document loaders and vector stores.

The second thing LangChain adds is an agent loop. LangChain describes an agent as a model that keeps using tools, round after round, until the job is finished. A tool is a function with clear inputs and outputs that the model is allowed to ask for. Tools let an agent fetch live data, run code, query databases and take actions. The model decides when to call a tool and which inputs to give it.

LangChain calls everything around those rounds the harness. That means the prompt, the list of tools and any middleware you add. Its main building block here is a function called create_agent. You build an agent from a model, a list of tools, a prompt and optional middleware. Middleware is optional extra code that shapes how the agent behaves. It can add guardrails (safety checks), retries and routing. When you turn a normal function into a tool, its docstring, the short note at the top of the function that explains it, becomes the tool’s description. Type hints, the labels that say what kind of value each input is, define what the tool accepts.

LangChain’s starter example builds an agent with one weather tool. The tool is a stand-in that always says it is sunny. You ask what the weather is in San Francisco. The model sees a weather tool and calls it with the city name. The harness runs the function and passes its output along. The model reads it and writes the final answer.

LangChain sits in a wider set of related projects. Underneath, LangChain agents run on LangGraph. LangGraph is a lower-level runtime: the engine that keeps long-running agents and workflows going. Through it, agents can save their progress, pause for a human to check a step, and survive failures by resuming where they left off. You do not need LangChain to use LangGraph. Deep Agents sits one layer above LangChain and adds planning, subagents and file system use. LangSmith is a separate platform for tracing (recording each step), evaluation and deployment. There you can inspect traces and tool calls.

A framework does not make the model any smarter. The docs say the model you choose directly shapes how reliable the agent is, because the model decides which tools to call and when to answer. The model also reads your tool descriptions to work out when to use each tool, so clear docstrings matter. Some workflows mix fixed, hand-written steps with steps the model decides. For those, the docs point to LangGraph rather than a plain LangChain agent.