Building a LangGraph ReAct Mini Agent

[@imadsaddik]

Oh Sam, thank you so so much. This solution is crazy simple. The agent never enters in loops unlike the solution I develop with pure langchain. The agent used to get stuck in loops when asking simple questions like hi or hello Thanks again

[@mohamedsamy9723]

HI sam , we hope to work more closely with open source models such llama3, command R some of this technique not working with such this models and need some of customisations we need to reuse this concepts with open source models and thanks a lot ❣️

[@viperactual]

Great stuff! Well done on the presentation.

[@parthdeshwal4419]

thanks man ...keep on the great work

[@shaunwhite8204]

Great Content! This was super helpful!

[@gauravgarg-wc4zl]

Great ! , i was able to swap the prompt and test it , great learning ...

[@deeplearning7097]

Very nice work, as always, thank you.

[@Diego_UG]

What a great explanation, thank you

[@satvikparamkusham7454]

This is very helpful!

[@pycoder8233]

Great video! I've been following your channel to learn LangGraph. I've watched your previous videos on LangGraph and have learned a lot. Could you please revisit one of your older projects using this new method? It would help us understand it better.

[@kenchang3456]

Thanks Sam.

[@juanmanuelzwiener4447]

Hi Sam, please provide a video about the SQL agent that manages fields with high cardinality to generate the correct query in Langgraph, improving even the ones uploaded in the oficial documentation

[@samwitteveenai]

do you have a dataset, I can't use client data for tutorials etc

[@juanmanuelzwiener4447]

@@samwitteveenai yes i do, want it to practice?

[@allok501]

thanks man, what about using this reasoner to supervise agents instead of tools?

[@samwitteveenai]

Good idea, I think I'll make something about controlling subgraphs with Reasoner.

[@ofrylivney367]

great content! thanks!!

[@MLGJuggernautgaming]

Great video! How could you have this print only the final response though from the stream? I can't figure it out. For example, I don't want it to print "First, i'll look up x..." just print the final response. I'm using astream_events, but can't see any way to descriminate the responses.

[@sergiozavota7099]

Thanks Sam, great video! One doubt I still have about ReAct agents in langgraph is the "component" that makes the LLM to think in the ReAct style. In the ReAct paper, the prompt instructs the LLM to think, act and observe. However in this example I don't see any of these instructions. Where are they "hidden" in this version with Langgraph?

[@samwitteveenai]

yeah I pointed this out in the video that they aren't using the standard prompts. You could put that in in there if you wanted. With Function Calling / Tool use and much better models we have today you don't need it as much, but for anything serious I would massively tune that prompt in the reasoner

[@johnpiscani5885]

Hi Sam, I am currently trying to follow this to learn more about this architecture. Right now, I am using a tool for getting a stock price and one for searching Wikipedia. When asking 'What is the stock price of Tesla and what can you tell me about the history of the company?', I keep running into a GraphRecursionError. I confirmed both tools work independently but when trying to put it together, I am seeing this recursion error, even with stock price and addition tools. Why may using two tools cause a recursion error?

[@autoflujo]

Nice tutorial as always Sam! Some months ago I tried to use this approach of one node for reasoning and one for tool calling but the LLM always went nuts. It iterates over and over with the same tool. Is there a way to control how many times the agent can call a tool or how many times it can iterate the tool node before saying “I don’t know”?

[@micbab-vg2mu]

thanks

[@srikanthch1524]

Very interesting and thanks for the code. I tried modifying your code to build code assistant agent. How to add human in the loop ?

[@RaviPrakash-dz9fm]

Does it internally make the LLM follow the Thought Action Observation pattern? Like, I don't see any thought part in the trace

[@ItachiUchiha-fo9zg]

I tried this with groq free api mixtral model... and i faced issues as the search tool didnt get the age sometimes.. so I have to change the prompt..."You are a helpful assistant tasked with using search and performing arithmetic on a set of inputs. you are only allowed to use the tools add, multiply, divide and search. repeat the tool until you think you got the expected output"

[@samwitteveenai]

yes certainly different models will respond to differently prompts. thanks for sharing that one

[@joaopaulopresa]

How to use it as a subgraph? Do I need a router? For example if I want to isolate heavy function calls from the main agent history? Is making it a tool an option? Can it have its own messages and state and be reset when it finishes processing like a subgraph?

[@samwitteveenai]

This is a really good idea. A number of people have asked for this, so I'll make something with showing off subgraphs and how you could use this kind of style for a subgraph.

[@waneyvin]

It is a better alternative than o1 model of openAI!

[@ynboxlive]

Thanks Sam! Can you do a comparison of this model to a multi-agent framework like crewAI?

[@samwitteveenai]

At a low level, CrewAI is using a lot of things like this from LangGraph and LangChain and other agent frameworks. The big difference is that CrewAI is not on a guided flow as strictly as something like LangGraph is.

[@arjunparmar6121]

what if you have math node with math tools and search tool separate as a node? you cant bind llm. how to approach this?

[@Lirim_K]

What happens if the history becomes so large that it falls outside the models context window?

[@samwitteveenai]

We can trim the history pretty easily, maybe I'll do a video about that soon.

[@Lirim_K]

@@samwitteveenai yes but then you’d lose valuable context and the model would “forget” stuff.

[@ibrahimhalouane8130]

I have been trying too many frameworks some of them looks over engineered including Langraph, you don't really have control on what's under the hood specially when you are using less capable models, I'm trying Phidata as a small library that just work yet flexible enough for most basic tasks. Can you make a video on how to use it to make an autonomous agents ?

[@MrAhsan99]

Can anyone share some langgraph projects which can help me land some job or freelance work?

[@Divyv520]

Hey Sam , very good video ! I was wondering if I can help you with more Quality Editing in your videos and make Highly Engaging Thumbnails which will help your videos to get more views and engagement . Please let me know what do you think ?

[@poisonza]

i might try to implement this without langraph... langchain shitty library and langraph is just a state transition library with some weird abstraction

[@samwitteveenai]

I don't totally disagree with you. I do have some agents vids coming with no frameworks used