Which Agentic AI Framework to Pick? LangGraph vs. CrewAI vs. AutoGen

[@nikosterizakis]

We spent 2 months in the summer of 2024 evaluating the same frameworks. More or less aligned to your findings. Our findings were published in our blog. However, in relation to your recommendations on closing, we would recommend LangGraph for anything complex, with Autogen sitting in the middle. Furthermore, the CrewAI LowCode studio is a subscription based model, not for free anymore, since their last major update (late Autumn 2024). The open source version you saw, I believe is the old version that is not supported anymore. It would appear that all 3 of the frameworks had major updates late 2024, with Autogen being still 'worked on'

[@WW_AI_Adventures]

@@nikosterizakis Thanks, I'm glad they are fairly aligned. I'm a big fan of langgraph personally, but it is hard to get started with so understand why people want something simplier to get started with. That's really frustrating, it seems to be a trend with agent building frameworks...

[@nikosterizakis]

@ we are a fan of it too. But, as you say too low level and difficult to grasp, especially for multi agent flows with conditional edges. The documentation was not brilliant at the start, but getting better. Their recent tutorials are a great improvement. The DeepLearning tutorial was a mess and potentially the perception of it being to make it work.

[@franknillard]

Amazing video Will! Looking forward to the consultancy!

[@WW_AI_Adventures]

@@franknillard Thanks - bring it on!

[@davidyoo647]

just subscribed, the video was very helpful in understanding the state of agent frameworks!

[@WW_AI_Adventures]

@@davidyoo647 Thanks 🙏

[@SyntharaPrime]

Great effort. Thank you.

[@rsrini7]

Thanks for the comparison. and its wonderful knowledge. is Pydantic AI Agent competes all of these ? Also, swam from OpenAI ?

[@WW_AI_Adventures]

@@rsrini7 swarm isn't a production ready system just yet so didn't include it. I'm not sure about PydanticAI just yet, but I plan to look into it!

[@ArjunSharma-lx7ev]

You’re doing a great job!

[@WW_AI_Adventures]

@@ArjunSharma-lx7ev thanks 🙏

[@jaggyjut]

Very insightful and learnt a lot from this video. Thank you

[@WW_AI_Adventures]

@@jaggyjut thanks!

[@peterm5039]

Thanks for the overview. My only issue is your rating for LangGraph on integrations seemed very cursory. You mention it integrates with LangChain and LangSmith and give it an 8. You seem to have graded the other two frameworks harder. As others have said, looking forward to see evals of other options like pydantic and semantic kernel.

[@WW_AI_Adventures]

@@peterm5039 Thanks for the feedback. My motivation for that is that because you can define nodes/tools/edges independently of any 'agent' abstraction, you could include any custom integration you needed, without having to interact with or update the modules code base.

[@arpeetgupta14]

Informative vedio 👍

[@camilocortes8074]

It would be interesting to see pydanticai in this comparation, nice video btw, incredile value!!

[@WW_AI_Adventures]

@@camilocortes8074 currently making this video!

[@AlbertoGarcia-on9kq]

Great video!

[@WW_AI_Adventures]

🙏 thanks

[@kovlabs]

Please review semantic kernel and pydantic ai

[@rajashekarakula3991]

Good high level comparison of different agentic frameworks, would have been better if gone bit deeper but good starting point, liked it!!

[@WW_AI_Adventures]

@@rajashekarakula3991 Thanks! If I do a follow up, what is it that you want to know?

[@DeepMimd007]

You have rated the frameworks based on their documentation, brilliant what could go wrong, everything they must have mentioned must be correct 👍

[@WW_AI_Adventures]

Their documentation, popularity, complexity (how many abstractions) and more... Saves you having to pick out the relevant bits!

[@DeepMimd007]

@ brilliant, why would any framework author/owner lie?

[@WW_AI_Adventures]

@DeepMimd007 Doesn't really matter if they lie or not, I care mostly about the structure of the framework and how sound it is of a concept. If the concept works and is sound, then as it's an open source project it will attract enough interest from good engineers. Also an open source project with such attention (see the stars at the start of the video), wouldn't get away with a blatant lie on their documentation. Also, quality projects often have quality documentation!

[@mohammedameen7409]

Great job 👏

[@WW_AI_Adventures]

Thanks! 🙏

[@ahmedelkamel147]

Good and in-depth comparison. What do you think about smolagents from Hugging face?

[@WW_AI_Adventures]

I like the fact that it integrates with HuggingFace models, so it could be useful for anything that you want to do quickly or for a proof of concept, as integrating with on device models can be a pain! It looks simple, so it depends what you want.

[@rousabout7578]

Just a heads-up, your video appears to be using Autogen 0.2. This month a released stable version, Autogen 0.4, has a brand-new architecture and drag and drop studio.

[@WW_AI_Adventures]

@@rousabout7578 Thanks

[@jorgebarrero5299]

I woul like if you can give some info about langflow

[@alx8439]

Thx, good comparison. But. It lacks the most important part - performance comparison. How good / bad these agentic frameworks solving the same set of tasks. With some statistical meaningful number of repetetions.

[@WW_AI_Adventures]

, I would like to do this, but I'm not sure how as some frameworks are more task based, others are more conversation focused, some are more research focused... Feels like comparing cars and vans and then saying vans are slower

[@alx8439]

@@WW_AI_Adventures google for GAIA (the benchmark). It's easy to implement something similar.

[@alx8439]

@WW_AI_Adventures have a look at GAIA benchmark

[@chowdarymygang]

Thanks for creating this video, its helpful. Can we create a RAG application with langchain/langgraph ? and when should i use langchain vs langgraph ? What do you think about Semantic kernel ?

[@WW_AI_Adventures]

@@chowdarymygang Thanks for getting in touch - I might make another of these for semantic kernel and smol agents

[@WW_AI_Adventures]

@@chowdarymygang yes you could use either for a rag application. I would use LangGraph the moment you want to add a 'loop' into the logic

[@chowdarymygang]

@@WW_AI_Adventures : thanks for the reply. Langgraph vs semantic kernel, which one would you pick for building an agent ? Any advantages of using langgraph instead of Semantic Kernel?

[@rishabhkhandelwal4627]

Also we have now two different version of autogen ag2 and autogen0.4, which one according to your analysis is better?

[@WW_AI_Adventures]

@@rishabhkhandelwal4627 at the moment the latter one (0.4) is only in preview. However they claim the more recent one is better at handling async requests - so I expect when it is officially released it will be better

[@markferris9632]

@@WW_AI_Adventures FYI 'AG2' is not (exactly) the same as 'AutoGen v0.2'. The original creators of AutoGen were from a university with a partnership with Microsoft. MS sort of took over ownership of that project's roadmap which didn't align with the creators' vision, so they branched off to create AG2. That's about all I know as I've not actually looked into the AG2 code examples yet, but did want to point out that AG2 is a different project and different team than AutoGen.

[@rthidden]

The CrewAI UI Studio is available with the red "Get Started" button under the image in this clip, kzbin.infoUgkxEMjpJbnzs--1ANh9-K1o2SE2INagLE15?si=bTCsx5yQQEuHJw5I

[@WW_AI_Adventures]

I tried that before and it was directing me somewhere that wasnt their studio... This time it worked though, so maybe some session funniness in the browser

[@ParthShukla-o3t]

Hi sir is there a way to connect to you ??

[@WW_AI_Adventures]

@@ParthShukla-o3t yes - check out the video description 👍👍

[@Jobeyhshxgs]

Pydantic ai works great with langchain.

[@WW_AI_Adventures]

Ah that's good to know 🙏

[@reserseAI]

phidata 😅🔥🔥🔥

[@WW_AI_Adventures]

@@reserseAI looks good tbh! Another good out of the box option. A templated API like this would be good if you could then build the templates with an orchestrator like LangGraph

[@HasnatAbdullahz]

smolagent where? 😁

[@WW_AI_Adventures]

@@HasnatAbdullahz I only picked a few of the top ones sorry!

[@andrew.derevo]

any one else feels like crewai is only a hype driven and almost useless in any production applications?

[@WW_AI_Adventures]

@@andrew.derevo I certainly got that feeling about Autogen. CrewAI looks ok as a framework but I think only if what you are trying to achieve is task based and well defined.

[@andrew.derevo]

@ we use langchain with lightllm as unified api proxy provider, works really good for now, langchain sometimes feels overloaded in some cases but definitely do the job. will check crewai again, probably something changed over time.

[@faridautomatic]

@@WW_AI_AdventuresCreeAI has limited documentation path to true production deployment at scale in the Cloud compared to LangGraph that offer multiple options: SaaS, Cloud (container) or even local dev! So for me CrewAI is the weakest in this area

[@orthodox_gentleman]

Absolutely agreed

[@alx8439]

I'd say even more broader. There's generally a lot of hype and self promotion in this agentic realm, not backed by any real-world performance tests.

[@alx8439]

There's also smolagents from huggingface

[@0730pleomax]

Pydantic AI is a strong competitor, have a check.

[@WW_AI_Adventures]

@@0730pleomax thanks - am in the process of making a video about PydanticAI, llama index workflows and smolagents & will release in a few days

[@mariuszsiera]

IMHO comparison is very shallow, it is based on documentation overview and not sure author has used all of them in practical problem solving. What would be more usual is comparing implementation of non trivial problem compare their implementation complexity, accuracy of result, performance and cost of changes.

[@WW_AI_Adventures]

@mariuszsiera thanks for the feedback. I have used 2/3 of them and wanted to see how the other shaped up. I considered something like that but as I was creating this video it felt like comparing apples to nuts. They are all quite different and have different problems that they would work better at solving.

[@roromemeqatar123]

Most of them will not survive it take years to mature currently waste time to choose any Bez we do not know which to win and adopted by industry

[@alexwoxst]

Pydantic AI is obviously the best

[@WW_AI_Adventures]

Oh yeah? I'm not so sure..

[@alexwoxst]

Haha maybe you’re right, but I find it nice. Not to many abstractions and easy to get insight with log fire. Have you tried it?

[@WW_AI_Adventures]

@alexwoxst Not yet - I had a look and it does look like a good API with simple abstractions, I will be taking a more detailed look though!

[@Jobeyhshxgs]

Pydantic ai is a revolution for me in making commercial grade ai apps. It's easy to swap out llms as requirements change so many other benefits

[@WW_AI_Adventures]

@Jobeyhshxgs Interesting, thanks for letting me know.

[@lenovo57787]

LangGraph documentation is complete and utter trash. More documentation does not mean better.

[@WW_AI_Adventures]

@@lenovo57787 have you seen the new course they have just released? It's pretty good. Their documentation has also been massively improved over the last 3 months or so, so I'd give it another go if it was a while ago

[@lenovo57787]

@@WW_AI_Adventures Yes, the LangGraph course is what I was mostly referring to. I'm on the last module of the course. It's not that new. It's at least 4 months old. Lance races through jupyter notebooks like he's giving a lunch and learn presentation to his colleagues at LangChain. Everything is overly complicated. Aside from the first module, every one after has been a pain to follow. Lance is just reading notebooks instead of teaching.

[@WW_AI_Adventures]

@lenovo57787 thanks for letting me know, I only followed the first module!