LangGraph: Multi-Agent Workflows

LangGraph makes it easy to construct multi-agent workflows, where each agent is a node, and the edges define how they communicate. In this video we will walk through three examples of multi-agent workflows
Python code: github.com/lan...
JavaScript code: github.com/lan...

[@christopherscheidel5431]

Fantastic work as always. Thanks to the LangChain team.

[@bqmac43]

These videos never cease to impress me. Straight-forward and effective. Thank you, LangChain team!

[@akashlives]

Thank you for constantly posting such videos. So easy to use and adopt.

[@Orcrambo]

Excited to get my hands dirty with langgraph. Hopping on now!

[@Sandheip]

LangGraph makes creating multi-agent processes easier. SymthOS is an essential viewing for anyone interested in cutting-edge AI frameworks. #AI #MultiAgentWorkflows #SymthOS

[@paraconscious790]

This is a super helpful series, a quick starter, easy to follow along with practical examples, thanks so many Harrison! Starting to experiment right away!

[@free_thinker4958]

That would be great if you add a disclaimer at the beginning of the video mentionning that it's for intermediate and advanced levels because beginners will feel lost and sometimes more confused just like me

[@ManiSaintVictor]

Thank you! Very helpful. I did the first 2 exercises. I'll be back for the 3rd.

[@ramzirebai3661]

This is incredibly beautiful

[@raphaelb7933]

Super helpful series. I delved a lot in langchain source code, docs to customize already well-established examples and started these series for LangGraph it's a great format thanks a lot.

[@tonyppe]

are you using ai to write youtube comments?

[@ilianos]

@@tonyppe just because you read the word "delve" ;-) ? look at the last words of his comment: missing punctuation marks :)

[@johnvicente796]

Great video. Can you PLS create an example in langgraph on how to use an SQL database tool and call the tool with an agent. More importantly do a RAG search with it. That would be helpful

[@glowstorm334]

For anyone facing any error , there are the two error I have faced : 1: Make sure Matplotib is installed in your environment 2: Change the name of Chart Generator to Chart_Genenrator , this fixes an error which the regex does not recognize the name of Chart Generator.

[@Mullheimer]

Great way of explaning. Thank you. Will dive into it sometime soon.

[@lucascresci509]

Bro is just a beast!

[@kevinkawchak]

Thank you for the discussion.

[@ZhambylYermagambet]

Awesome! Thank you guys for great work

[@dvirbenor2690]

Very interesting and well explained! Thanks 👌

[@drtristanbehrens]

Such a cool video!

[@adandrea]

Great video. One bit things I'm trying to figure out: how can the tools access the graph state? I mean custom tools. I really need them to have context of the conversation, user session metadata, etc, and I can't past them that info if the tools are called just with the parameters filled with syntetic data generated by the llm. I'm sure there is a way to do it but can't seem to fiture it out. Thanks!

[@aryangautam7161]

I tried to implement multiagent using supervisor agent and running into issue where it couldn't parse function call. langchain_core.exceptions.OutputParserException: Could not parse function call: 'function_call

[@rahulvb5044]

In the supervisor example , the model returns , { "function_call" : { "arguments" : "{"next","Coder"}", "name" : "route" } How is this used to determine the next agent. Or how is Agentstate populated with next : Coder after getting the above out put from model. I can see a JsonOutputFunctionsParser. But i cant understand, how next value is determined from that

[@HiKami-f5o]

In conditional edges setup. You parse it to dict and then make a conditional navigation depending on 'next' value

[@octopusfinds]

Thank you!

[@estebanmunoz1637]

I can't make this example work using AzureOpenAI, I am receiving the error: 'create() got an unexpected keyword argument 'functions'' after executing: result = agent.invoke({"input": "what's the weather in SF?", "intermediate_steps": []}) Are Agents supported using AzureOpenAI? I have been trying different formulas but I cannot make my agents work using AzureOpenAI

[@jimjimjimjimjim]

Does anyone get issues where the supervisor agent loops over and over calling the sub agents?

[@calebadobah641]

Yes same. Have you managed to fix it?

[@aryangautam7161]

Are there examples on how you can use multi agent workflow that doesn’t involve openAI function calling?

[@waneyvin]

great job! please advise how if I wanna add reward or policy into the state for further decision making sake.

[@ragsAI]

Hi! Is Langchain integratable/compatible with redshift/databricks? (especially the text-to-sql framework)? Thank you.

[@adityagaurav2816]

I tried to implement supervisor based multi agent framework for my use case but after my workers return something to supervisor , supervisor doesn't call FINISH and get into a loop of calling the same worker again and again. Has anyone faced this issue ? or know how to fix this ?

[@zelx7567]

Thank you langchain team, this helps a lot. In the example multi-agent graph all agents share the same instance of a large language model (LLM), is it possible to use different LLMs for different agents?

[@StaraX95]

yep

[@mertsengil4697]

u can use llama3.1 with groq

[@VijayasarathyMuthu]

Could you show how to bring Humans in the Loop in each architecture or any one of them?

[@joehenri1]

What is the differences between CrewAI and LangGraph ?

[@yuripourre]

CrewAI is a thin wrapper around langgraph.

[@mertsengil4697]

Hello, thank you for the vide. I have a problem. I created my own tools instead of "tavily search" and "PythonREPL" tools. I did everything same with your code. But i cant get the end token {'supervisor' : {'next' : 'FINISH'}}. why? can anyone help me pls?

[@mertsengil4697]

Continuous 1. tool... 2. tool... 1. tool. ... As the 2nd tool, it loops through my tools, but it never goes to the finish token and finishes the loop.

[@陈晓耀]

not work with the current version of langgraph/langhchain, issues never solved

[@googleyoutubechannel8554]

Oh no, I feel so sorry for everyone trying to use this, you don't have nearly enough basic logging functionality or debugging / investigation tooling in langchain to make using multiple prompt stages remotely tolerable, I wonder how long it will take people trying to actually implement this to solve a problem to figure it out.

[@ste7081]

didn’t they make langsmith exactly for this?

[@xfldiazx]

​@@ste7081 but it's not private :( honestly a framework should not rely on a paid service for basic functionality, I wish I could use it though

[@Stormyys]

How do we get access to langsmith?

[@LangChain]

DM Harrison on twitter @hwchase17 :)

[@carterjames199]

How can I request access to langsmith??

[@LangChain]

hey, if you drop Harrison a message on twitter @hwchase17 he'll get you access

[@carterjames199]

@@LangChain already done he was very fast to reply

[@manfyegoh]

hi how do i get access to langsmith?

[@LangChain]

Drop Harrison a message on twitter @hwchase17 and he'll sort you out

[@manfyegoh]

@@LangChain thanks! i just gotten my access!

[@Kenykore]

Need access to langsmith

[@LangChain]

DM Harrison on twitter @hwchase17 :)

[@adamwilfrid3271]

requesting for Langsmith access please 😊

[@LangChain]

DM Harrison on twitter for access @hwchase17 :)

[@IAMTHEMUSK]

The last one is far from easy to implement. I wonder if learning langgraph is worthy instead of learning how to do the same from scratch. It feels like learning a new programming language.

[@RonnyAndersson-p6z]

Great video! Can you hook me up with Langsmith?

[@RonnyAndersson-l2q]

Great video! Can you hook me up with langsmith? :)

[@LangChain]

Shoot Harrison a message on twitter @hwchase17 :)

[@avi7278]

What's your Twitter?