langgraph

An open-source AI framework for building stateful, multi-actor applications with large language models (LLMs).

LangGraph

Hello everyone. Today I want to talk about LangGraph, a new library that we're releasing. So LangGraph builds on top of Lang chain and makes it really easy to create agents and agent run times. 

So what exactly is an agent and an agent runtime? So in Lang chain, we define an agent as a system powered by a language model that decides what action to take. There's then an agent runtime that basically

Runs that agent in a loop. Calls that agent. Decides what action to take. Then takes that action and then get records the observation. Then passes that back in and starts to loop over again. Continues going through that loop until the agent decides that it is finished. 

So we've made it easy over the past few months to customize agents in Lang chain with Lang chain expression language.

I'd recommend checking those out if you haven't. What we're doing with LangGraph is we're making it easy to customize the agent runtime. So previously the agent runtime was always the agent EX class. This was basically something that ran in a loop and called uh Tools in a specific way and handled errors in a specific way. That was great but it was just one way to create this.

Runtime and so we want to create more ways to create this runtime and ways to create them more flexibly and dynamically. A key part of this runtime is the ability to add cycles. So you know the whole point of the agent runtime is running this agent, this LLM powered agent in a loop. And so you need to be able to have cycles and Lang chain expression language and other, uh you know, DAG likee.

Frameworks are non they're not cyclical. Um and so that's what we're introducing L graph for. Um it's it's a way to create these cyclical agent run times. So that's hopefully an overview of what Lang graph is and and why we created it. 

The rest of the videos in this series are going to focus on two of the main agent run times that we've added to start. So first we've added a agent executor very.

Similar to the agent executor in uh LangChain. So we've recreated that agent executor with LangGraph. And then we've added a chat agent executor. The chat agent executive takes in a list of messages and then represents the agent State just as a list of messages. And so it returns a list of messages as well. The reason we did this is that a lot of the newer models are chat-based models and they.

Intrinsically represent function calling um as as basically parameters of part of a message. And then they also have uh function responses being a separate type of message. And so representing this agent state as a list of messages is very natural for these types of models. 

And so we've added a separate agent executor particularly for that. Then I'm going to highlight a bunch of different ways.

That you can modify the base agent executors to do things like add human in the loops. Force specific tool to be called first and other exciting things like that. Hope you enjoy.

Note, that there may be slight differences in text and video. Watch video for the original content.

LangGraph Video