I'm building something BIG behind the scenes, launching soon! ottomator.ai Come join the oTTomator Think Tank - home of the oTToDev community! thinktank.ottomator.ai
why nobody works on a grounding agent, something that goes to the docs online or local docs and validates the AI suggestions? doesn't matter through which tooling, but this would be the best tool, which can be a proxy between AI responses to the created diffs in the IDE for the autocomplete or for the chat and other types of interaction. Cursor tried to make it but they load docs into a RAG so your "@docs" tag will only grab a small chunk of the referenced documentation from RAG and pass to context, but that's not enough, something has to go during every edit suggestion over the suggested code and validate it all against the latest versions of documentation of whatever tools and libs are used in the suggested code snippet. would definitely help if it also allows to iterate for the AI behind the scenes based on the found mismatches with docs. What do you think, does it make sense to develop something like that? (and why nobody does it?)
@@babyfox205from what I read about goggles code assist it has doc knowledge, they talk about collaborating with lib maintainers. To me it makes a ton of sense that the assistant would just read dependency versions and get the right docs by itself, then somehow use them to steer and maybe check output as you're saying.
Really excited about PydanticAI. Having built some swarm agents for a personal asssitatnt, this library adds all the missing pieces and make for a complete but simple framework without all the not-needed abstraction like crewai, autogen, langchain etc. Please continue with more videos on this library. :)
@DouhaveaBugatti Let me guess... you use your mom's gpt account. Swarm is a experimental framework released by OAI 2 months ago. Agency Swarm is something different.
I like the 3 word description of what an agent is: “pipes, memory, tools”. I can’t see any support for multiple agent flows. I assume the answer is “you wire them together with python” like with the atomic agents framework…?
@@ColeMedin I'm trying to adjust your code to use a local model like vanilj/Phi-4:latest but getting some errors. If I figure them out I'll share here. Thanks for your video :)
OK #1 I needed to add `from pydantic_ai.models.ollama import OllamaModel` and change the ternary model assignment to `... else OllamaModel(llm)` and #2 choose a different ollama model that actually supports tools (I picked granite3.1-dense:latest) I have a feeling I'll be watching more of your content to have more fun with this - thanks!
Is there any way to integrate it within n8n. I am not a coder unfortunately but I wanna put my best foot forward and integrate as many important things as possible
Great question! There are a ton of ways Pydantic AI can be integrated with n8n! I'm actually going to be making content on this in the near future because I love both platforms a ton. My favorite way to use n8n with frameworks like this is to use n8n for my agent tools (interacting with third party services like Slack and Google Drive) and then using the framework for all the custom agent logic.
N8N is good for hobbyists but when it comes to building scalable applications , it’s not that great. If you still want to use N8N, you will have to work with worker nodes, Railway has pre-built N8N nodes with redis.
I want to build AI agent that can analyze a logistic data in excel, discover patterns and plan logistics or do something like conditional filter in the excel data to create a trucking plan. Can i build that with it?
Yes that is correct! Though I prefer Pydantic AI over Langchain just because Langchain has become very bloated and isn't considered as production ready even by Harrison and Sam (some of the original creators of Langchain).
@@ColeMedin thanks, this clears my confusion. Do you have plan to update your AI Agents Masterclass to use Pydantic AI instead? Personally, I also feel that Langchain gives too much abstraction, and yeah it's quite bloated by now
I'm actually putting out a AI agent series on my channel currently (already have a few videos out) where I'm using Pydantic AI! It's definitely less introductory than my AI Agents Masterclass though so yeah I am planning on doing a big revamp using Pydantic AI instead of LangChain (though still incorporating LangGraph because LangGraph is incredible).
Are you going to be building more agents in the future? Also can this work with your n8n, flowise and openwebui config. Local ai packaged? Or are they two totally different things? Do I sack off the n8n stuff now then?
Great questions! - Yes, I'm going to be building more agent use cases with Pydantic AI in the near future! - Yes, you could create an Open WebUI pipeline to use Pydantic AI agents, and have these agents leverage n8n for tools just like I did with Flowise! I wouldn't use both Flowise and Pydantic AI at the same time, though I would recommend prototyping your agents with Flowise initially and then moving them to Pydantic to make them more production ready! I am even thinking of adding some Pydantic AI templates into the local AI starter package. - Don't sack off n8n! It's still the best no/low code AI automation platform in my mind, and I use it to create the tools for my agents all the time regardless of what framework I am using (LangChain, Pydantic AI, Flowise, etc.)
@@ColeMedinPerhaps you can bring up a few of these vital questions in your future videos to address them. Might help tons of people. Glad I read this. Tnx
@@ColeMedin Kinda reminds me of the Ansible/Terraform debates in DevOps circles... There's a lot of overlap between the two tools, but their strengths are rather different, so they are best used together.
Good question! This is actually not going to replace n8n at all! It's still much easier to use n8n to set up third party integrations with services like Google Drive or Slack compared to doing it in custom code with a framework like Pydantic AI. So what I actually really like doing is using n8n to create my agent tools, and then making the actual agent in a Python framework like Pydantic AI/LangChain/Crew AI.
Great question! Pydantic AI is a "foundational" agent framework, it gives you a baseline to build literally any kind of agent you want. Web research agents, agents to manage your email/messages, RAG agents to answer questions over your knowledgebase, the list goes on!
Found other video about ForestSwarm maybe its based on Swarm framework, not trying it yet but they claim it to be production grade, wonder how about if compare it with pydantic
You bet! Streamlit and Windsurf are two very different things. Streamlit is a Python UI library that makes it super easy to build web interfaces. Windsurf on the other hand is an AI coding assistant IDE. In fact you can use Windsurf to build your Streamlit interfaces - I do that a lot!
@@ColeMedin Thanks man, I bought windsurf and doing awesome things , learning on the way backend :) I will try to build some streamlit and check it out :) by the way in windsurf i get a lot of failed to run command rmdir: exec: "rmdir": executable file not found in %PATH% Do you might know how to fix this and add to path?
Good question! There are a ton of use cases for this and a lot of it depends on the business needs. Some examples are - transcribing and taking notes on meetings talking about sensitive info, agents that leverage fine tuned LLMs, RAG agents with knowledgebases filled with intellectual property, etc.
Thank you and good question! You can override the base URL of the AsyncOpenAI instance just like what I did with Ollama in the video! console.groq.com/docs/openai
Cool, I was looking forward to this one. I have a question about uncensored models, are there any that are really 100% uncensored, and could you attach a web search tool to any of those? I just think it would be ideal to have a model like that. Sometimes I ask political questions and I feel like the AI might just be giving me a biased reply.
I'm glad! Good question - I have heard that Dolphin is the most uncensored model out there from a lot of people's experience. I would give it a shot! ollama.com/library/dolphin-llama3
Good question! Certainly not in my mind, n8n is still the easier platform to build prototypes with and integrate with different apps. I actually love using them together - Pydantic AI for making the agents and n8n for the agent tools!
It has been a while since I've used CrewAI, but I have been keeping up with it a bit! You are certainly right that it is easier to build agents with CrewAI compared to Pydantic AI. Where Pydantic AI shines is with some of the deeper stuff to make your agents more production grade like testing, context management, LLM output validation, logging/monitoring, etc. CrewAI has features for a lot of these things, but certainly not as developed in my mind. For example, if you check out the testing page in the CrewAI documentation: docs.crewai.com/concepts/testing It feels very barebones compared to the Pydantic AI documentation for testing: ai.pydantic.dev/testing-evals To me it feels CrewAI's goal is more around abstracting things to make building agents easy. And that is important and it's a great platform for that. But Pydantic takes the "grade" of your agents to the next level.
After watching this I am more convinced that SpringAI Java is way ahead in things that will be important going forward - type checking, observability, production readiness etc. We are anyway calling an API for the AI stuff, the remaining stuff is plain old software and Java has no match when it comes to observability, testing, debugging and general idiot proofing. Spring AI already has all the things you mentioned and much more...actually these things are basic hygiene when it comes to any Java/Spring application.
You know I don't like Java personally, but I definitely think you could be on to something! I certainly respect the language and frameworks like SpringAI. Have you developed much with it yourself?
Good to know that Java is not lagging behind - I actually thought it was xD. I am a dotnet C# dev playing with semantic kernel. I will need to check this pydanticAI out in python to see how much overlap there is. Exciting stuff, but it seems to me that those AI agent frameworks are popping up just like JS web frameworks. At the end of the day we have to choose something! :)
Yes I have! And I do like it a lot and I love VRSEN's content too! Pydantic AI offers a lot more as far as taking your agents to production level, though. A lot of the features I talk about in the video are missing from Agency Swarm just like a lot of other frameworks like Crew AI, LangChain, Phidata, etc. It is a great platform to build agents quickly though, especially for prototyping! It's certainly a bit easier to get started with compared to Pydantic AI since it is more abstracted.
Good question! I would compare Pydantic AI to LangChain but not LangGraph. Pydantic AI and LangChain are both agent frameworks to help you build agents from the ground up. LangGraph on the other hand is a framework for orchestrating agents in workflows. So Pydantic AI + LangGraph is actually a really powerful combo I will be creating content on in the near future!
That's fair, and honestly what I thought as well at first! But once I really dove into the features it became clear it offers a lot more than other frameworks. Not necessarily for getting started building agents, but with actually taking them to production level. I've built a TON of agents with frameworks like LangChain, and while these frameworks are great, they're missing a lot of what I cover in this video!
@@ColeMedinespecially cause you know a lot of frameworks, have programming understanding and also know how to set a proper system prompt I wonder why you say this is THE framework. May I missed the point in the video, but I think I haven't seen any real difference? 🥲
@@1brokkolibaum Good question! The main features I listed out in the video - context management, logging/monitoring, testing/evaluation capabilities, error handling, and LLM output validation are all done very nicely with Pydantic AI that are often times missed or messy in other frameworks. All of these features are more important once you are past the prototyping phase for your agents, so initially it might not feel super different than other frameworks. But once you try to really get your agent ready for production, this framework really gives you what you need! I hope that makes sense!
@@1brokkolibaum To add to the "why". Context Management: Maintains coherent interactions across multi-step tasks. Example: A customer support chatbot remembers a user’s issue across multiple messages to provide a complete resolution without asking repetitive questions. Logging/Monitoring: Tracks behavior for debugging, insights, and real-time issue detection. Example: Logs show an e-commerce agent’s recommendations, allowing developers to identify and fix why irrelevant products were suggested. Testing/Evaluation: Validates performance across scenarios and ensures changes don’t break functionality. Example: Simulate scenarios where a finance bot calculates mortgage rates to ensure accurate responses even for edge cases like zero interest or maximum loan terms. Error Handling: Prevents total failure with robust fallback mechanisms. Example: A task automation agent encounters invalid input and gracefully alerts the user instead of crashing, offering guidance for correction. LLM Output Validation: Ensures responses meet required standards, reducing unpredictability. Example: A legal document drafting bot ensures the output follows proper formatting, avoids factual inaccuracies, and adheres to specified legal templates.
It is! Though Flowise doesn't have all these production grade features that Pydantic AI does. It's great for prototyping though - I love using it for that and that's why I made a video on Flowise recently! I'm sure Langfuse helps with some of what is missing from Flowise, but I'm thinking it wouldn't quite be enough.
Why show off with developing new code, when you modified their example? If you just modify their example code, It would be more useful, easier to remember, comprehend ...
I actually very much took my own spin on this example - I only used their example as a base but I changed a lot and added the whole web research part of the agent, running with Ollama, and the Streamlit UI!
The best ai builder out there is lovable ai by far better than bolt I'm using it right now to build a spotify mixed with iTunes site what's your go to ai builder😊
Yeah I have heard great things about Lovable! These AI coding assistants are quite different from agent frameworks like Pydantic AI though, one is using AI to code, the other is a framework to help you create agents.
Thank you! Honestly I stick to what I said in the title, I do think that Pydantic AI offers a way to build agents that surpasses all other frameworks. Not necessarily in getting started with building agents, but more making them production grade so you can actually build a product around them!
I'm building something BIG behind the scenes, launching soon!
ottomator.ai
Come join the oTTomator Think Tank - home of the oTToDev community!
thinktank.ottomator.ai
Theres a tiny bug below your website wpforms id='21373' title='false' description='false'] is available in plain
[wpforms id='21373' title='false' description='false'] looks like a bug
This seems like a worse version of CrewAI. Have you used CrewAI lately? Thoughts?
why nobody works on a grounding agent, something that goes to the docs online or local docs and validates the AI suggestions? doesn't matter through which tooling, but this would be the best tool, which can be a proxy between AI responses to the created diffs in the IDE for the autocomplete or for the chat and other types of interaction. Cursor tried to make it but they load docs into a RAG so your "@docs" tag will only grab a small chunk of the referenced documentation from RAG and pass to context, but that's not enough, something has to go during every edit suggestion over the suggested code and validate it all against the latest versions of documentation of whatever tools and libs are used in the suggested code snippet. would definitely help if it also allows to iterate for the AI behind the scenes based on the found mismatches with docs. What do you think, does it make sense to develop something like that? (and why nobody does it?)
@@babyfox205from what I read about goggles code assist it has doc knowledge, they talk about collaborating with lib maintainers.
To me it makes a ton of sense that the assistant would just read dependency versions and get the right docs by itself, then somehow use them to steer and maybe check output as you're saying.
My god man you are a full time job just to keep up ... keep going .. love the material and the examples.
Haha I appreciate it! Thank you so much for your support and the kind words!
Really excited about PydanticAI. Having built some swarm agents for a personal asssitatnt, this library adds all the missing pieces and make for a complete but simple framework without all the not-needed abstraction like crewai, autogen, langchain etc. Please continue with more videos on this library. :)
I'm glad you're excited too! Yeah I'm certainly going to be making more content on Pydantic AI!
@@ColeMedin
Hey it's similar to agency swarm
Which came more than a year before it.
😂
@DouhaveaBugatti Let me guess... you use your mom's gpt account. Swarm is a experimental framework released by OAI 2 months ago. Agency Swarm is something different.
Really happy to know I almost understand what you’re talking about. I’m almost there
We all need a little validation.
All of these AI videos are basically vocabulary soup
Haha I'm glad! Let me know what isn't crystal clear though because I would love to create more content around making that straightforward!
@@mikeantrI see what you did there 😂
You have a great way of explaining subjects! Glad to see your enthusiasm for pydanticAI 😁
Thank you very much Andy - I appreciate it a lot!
Wow, nice explanation about Pydantic AI and production ready agents. You have a talent to put this things. Thank You!
Amazing, thanks for the deep dive! nice to see internals of how this differs from the other frameworks
Alhamdulila!
This is awesome Cole! I have been thinking of building an app which integrates weather.
I like the 3 word description of what an agent is: “pipes, memory, tools”. I can’t see any support for multiple agent flows. I assume the answer is “you wire them together with python” like with the atomic agents framework…?
Yes that is correct! If you want agents to invoke other agents, you just create the agents as another "tool" that is callable by other agents.
Thanks for this. I was wondering which framework to tackle and learn and this video convinced me to go with Pydantic AI
Awesome to hear! I hope it helps you build some incredible agents!
Awesome sauce! Great video. Looking forward to testing this
Thanks, I'm glad you liked it!
Appreciate the video. Awesome knowledge. Can you please make sure the IDE is clear? It's hard to make out what you're actually doing.
You bet, thank you! Yeah in future videos I'll be sure that I'm zoomed into the IDE!
thanks for the great content dude
You bet man!
Great dome let’s do more agent use cases with pedantic and extend its api integration capabilities
Does pydantic AI enable a code interpreter/code execution workflow? Code Generation --> code execution --> feedback/retry
Yes you could make this happen pretty easily with Pydantic AI! I would actually suggest using it along with LangGraph for that kind of workflow.
Could you please make video about what is the system configuration required for running llm locally, it is very slow when I run in my normal machine
I am planning this actually, probably for next month! Which LLMs are you trying to run and what are your specs?
@@ColeMedin I'm trying to adjust your code to use a local model like vanilj/Phi-4:latest but getting some errors. If I figure them out I'll share here. Thanks for your video :)
OK #1 I needed to add `from pydantic_ai.models.ollama import OllamaModel` and change the ternary model assignment to `... else OllamaModel(llm)` and #2 choose a different ollama model that actually supports tools (I picked granite3.1-dense:latest)
I have a feeling I'll be watching more of your content to have more fun with this - thanks!
Do you know if you can integrate it with RAG?
Yes you sure can, and I will be making a video on exactly this (Pydantic AI with RAG) in the near future!
Is there any way to integrate it within n8n. I am not a coder unfortunately but I wanna put my best foot forward and integrate as many important things as possible
Great question! There are a ton of ways Pydantic AI can be integrated with n8n! I'm actually going to be making content on this in the near future because I love both platforms a ton.
My favorite way to use n8n with frameworks like this is to use n8n for my agent tools (interacting with third party services like Slack and Google Drive) and then using the framework for all the custom agent logic.
N8N is good for hobbyists but when it comes to building scalable applications , it’s not that great. If you still want to use N8N, you will have to work with worker nodes, Railway has pre-built N8N nodes with redis.
I want to build AI agent that can analyze a logistic data in excel, discover patterns and plan logistics or do something like conditional filter in the excel data to create a trucking plan. Can i build that with it?
Yes you 1000% can!
Will like to reach out
PydanticAI is awesome. Simple. Langchain etc are over complicated. Already built some useful things with this. Great video btw.
Thank you and yeah I feel the same way!
is PydanticAI in the same category with Langchain? I mean they both serve the same purpose and I can choose any of them if I want to do the same task?
Yes that is correct! Though I prefer Pydantic AI over Langchain just because Langchain has become very bloated and isn't considered as production ready even by Harrison and Sam (some of the original creators of Langchain).
@@ColeMedin thanks, this clears my confusion. Do you have plan to update your AI Agents Masterclass to use Pydantic AI instead? Personally, I also feel that Langchain gives too much abstraction, and yeah it's quite bloated by now
I'm actually putting out a AI agent series on my channel currently (already have a few videos out) where I'm using Pydantic AI! It's definitely less introductory than my AI Agents Masterclass though so yeah I am planning on doing a big revamp using Pydantic AI instead of LangChain (though still incorporating LangGraph because LangGraph is incredible).
Are you going to be building more agents in the future? Also can this work with your n8n, flowise and openwebui config. Local ai packaged? Or are they two totally different things? Do I sack off the n8n stuff now then?
Great questions!
- Yes, I'm going to be building more agent use cases with Pydantic AI in the near future!
- Yes, you could create an Open WebUI pipeline to use Pydantic AI agents, and have these agents leverage n8n for tools just like I did with Flowise! I wouldn't use both Flowise and Pydantic AI at the same time, though I would recommend prototyping your agents with Flowise initially and then moving them to Pydantic to make them more production ready! I am even thinking of adding some Pydantic AI templates into the local AI starter package.
- Don't sack off n8n! It's still the best no/low code AI automation platform in my mind, and I use it to create the tools for my agents all the time regardless of what framework I am using (LangChain, Pydantic AI, Flowise, etc.)
@@ColeMedin 🐐
@@ColeMedinPerhaps you can bring up a few of these vital questions in your future videos to address them. Might help tons of people. Glad I read this. Tnx
@@ColeMedin Kinda reminds me of the Ansible/Terraform debates in DevOps circles... There's a lot of overlap between the two tools, but their strengths are rather different, so they are best used together.
Awesome bro, exactly what I’m looking for! Thx
So this gonna replace n8n ? Like i know n8n is for automation but rven this also for automation right
Good question! This is actually not going to replace n8n at all! It's still much easier to use n8n to set up third party integrations with services like Google Drive or Slack compared to doing it in custom code with a framework like Pydantic AI.
So what I actually really like doing is using n8n to create my agent tools, and then making the actual agent in a Python framework like Pydantic AI/LangChain/Crew AI.
@@ColeMedin so basically n8n is still the king for automaiton and pydantic ai is subtitute for LLM creation tools like flowise etc ?
great content !! glad I found your channel
What are the potential use cases of developing with Pydantic AI?
Great question! Pydantic AI is a "foundational" agent framework, it gives you a baseline to build literally any kind of agent you want. Web research agents, agents to manage your email/messages, RAG agents to answer questions over your knowledgebase, the list goes on!
Aweasome video! which one do you recommend learn first? pydantic ai or langchain
Thank you! I'd recommend starting with Pydantic AI - definitely easier to get into it!
Found other video about ForestSwarm maybe its based on Swarm framework, not trying it yet but they claim it to be production grade, wonder how about if compare it with pydantic
Interesting! I haven't heard of ForestSwarm yet so I'll have to try it out and compare!
Downside of Pydantic at the moment is no Anthropic / Sonnet support.
Yeah I agree! I hope that add it really soon.
I might have to build this into my ComfyUI slash Ollama art agent project
Is that possible?
thanks! What is the benefit of using streamlit vs windsurf?
You bet! Streamlit and Windsurf are two very different things. Streamlit is a Python UI library that makes it super easy to build web interfaces. Windsurf on the other hand is an AI coding assistant IDE. In fact you can use Windsurf to build your Streamlit interfaces - I do that a lot!
@@ColeMedin Thanks man, I bought windsurf and doing awesome things , learning on the way backend :)
I will try to build some streamlit and check it out :) by the way in windsurf i get a lot of failed to run command rmdir: exec: "rmdir": executable file not found in %PATH%
Do you might know how to fix this and add to path?
Great video! Excited about the production quality tools from pydantic ai. Does it play nicely with LangGraph?
Thank you! Yes LangGraph and Pydantic AI is a GREAT combo and I'll be making content on it soon actually!
I don’t know how anyone learned to code before LLMs. I still don’t know how even with them
Are you working fulltime on your projects? Amazing dude anyway ❤
Thank you very much! Yes - my UA-cam channel and the platform I'm building behind it (oTTomator.ai) is what I am doing full time!
@@ColeMedin wishing you all the best bro. Much love from Ghana.
I'm definitely joining this community.
Newbie here, could you make a full tutorial on how to build AI agent with PydanticAI? 🙏
I sure will be in the near future!
It's strange that you are unable using ollama streaming output with pydantic ai , because I could do that with fastapi and ollama.
Yeah I've been able to do that too! Not really sure why it's different here unfortunately
I appreciate your hard work.
Which software did you use for recording.
Thank you! I use OBS!
What workflows would you only use on-prem / local models for?
Good question! There are a ton of use cases for this and a lot of it depends on the business needs. Some examples are - transcribing and taking notes on meetings talking about sensitive info, agents that leverage fine tuned LLMs, RAG agents with knowledgebases filled with intellectual property, etc.
Cheers Cole. I love your content.
Who else would you say is producing content on the same niche you are excelling on?
Hey Cole! Really liked the video.
I have a question. How would you use Groq's API instead of OpenAI's in Pydantic?
Thank you and good question!
You can override the base URL of the AsyncOpenAI instance just like what I did with Ollama in the video!
console.groq.com/docs/openai
Cool, I was looking forward to this one. I have a question about uncensored models, are there any that are really 100% uncensored, and could you attach a web search tool to any of those? I just think it would be ideal to have a model like that. Sometimes I ask political questions and I feel like the AI might just be giving me a biased reply.
I'm glad!
Good question - I have heard that Dolphin is the most uncensored model out there from a lot of people's experience. I would give it a shot!
ollama.com/library/dolphin-llama3
What about flowise ?
you're the GOAT! thanks man.
You bet man! :D
Sooo..... I guess this is the answer on how to get "Run agents in the backend as opposed to a single model call" into oTToDev then...? 😅
So this over N8N?
Good question! Certainly not in my mind, n8n is still the easier platform to build prototypes with and integrate with different apps. I actually love using them together - Pydantic AI for making the agents and n8n for the agent tools!
When was the last time you used CrewAI? I would say that it is much easier to use that’s PyndanticAI. It has all the same functionalities and more.
It has been a while since I've used CrewAI, but I have been keeping up with it a bit!
You are certainly right that it is easier to build agents with CrewAI compared to Pydantic AI. Where Pydantic AI shines is with some of the deeper stuff to make your agents more production grade like testing, context management, LLM output validation, logging/monitoring, etc.
CrewAI has features for a lot of these things, but certainly not as developed in my mind. For example, if you check out the testing page in the CrewAI documentation:
docs.crewai.com/concepts/testing
It feels very barebones compared to the Pydantic AI documentation for testing:
ai.pydantic.dev/testing-evals
To me it feels CrewAI's goal is more around abstracting things to make building agents easy. And that is important and it's a great platform for that. But Pydantic takes the "grade" of your agents to the next level.
Is it possible to add files?
Could you clarify what you mean by that?
After watching this I am more convinced that SpringAI Java is way ahead in things that will be important going forward - type checking, observability, production readiness etc. We are anyway calling an API for the AI stuff, the remaining stuff is plain old software and Java has no match when it comes to observability, testing, debugging and general idiot proofing. Spring AI already has all the things you mentioned and much more...actually these things are basic hygiene when it comes to any Java/Spring application.
You know I don't like Java personally, but I definitely think you could be on to something! I certainly respect the language and frameworks like SpringAI. Have you developed much with it yourself?
Curious here myself???
Good to know that Java is not lagging behind - I actually thought it was xD. I am a dotnet C# dev playing with semantic kernel. I will need to check this pydanticAI out in python to see how much overlap there is. Exciting stuff, but it seems to me that those AI agent frameworks are popping up just like JS web frameworks. At the end of the day we have to choose something! :)
Nice content, thank you
Thank you - you bet!
Someone could make a killing putting a GUI wrapper around this to no-code it.
Yeah no kidding!!
Have you tried agency swarm?
Yes I have! And I do like it a lot and I love VRSEN's content too!
Pydantic AI offers a lot more as far as taking your agents to production level, though. A lot of the features I talk about in the video are missing from Agency Swarm just like a lot of other frameworks like Crew AI, LangChain, Phidata, etc.
It is a great platform to build agents quickly though, especially for prototyping! It's certainly a bit easier to get started with compared to Pydantic AI since it is more abstracted.
Do you think this is better than Langgraph?, I actually dislike langchain but I like Langgraph.
Good question! I would compare Pydantic AI to LangChain but not LangGraph. Pydantic AI and LangChain are both agent frameworks to help you build agents from the ground up. LangGraph on the other hand is a framework for orchestrating agents in workflows. So Pydantic AI + LangGraph is actually a really powerful combo I will be creating content on in the near future!
I need your desk
Hopefully people understand what it truly means 👍 big step forward
It sure is! I appreciate you saying that Mike :)
perplexity api for the win
I've heard good things for sure! You think it's better than the Brave API for building web research agents?
Interesting 🤔 but I am not sure if this is really different from what we already have. Doesn't look different to some other agent frameworks 🤔
That's fair, and honestly what I thought as well at first!
But once I really dove into the features it became clear it offers a lot more than other frameworks. Not necessarily for getting started building agents, but with actually taking them to production level.
I've built a TON of agents with frameworks like LangChain, and while these frameworks are great, they're missing a lot of what I cover in this video!
@@ColeMedinespecially cause you know a lot of frameworks, have programming understanding and also know how to set a proper system prompt I wonder why you say this is THE framework. May I missed the point in the video, but I think I haven't seen any real difference? 🥲
@@1brokkolibaum Good question! The main features I listed out in the video - context management, logging/monitoring, testing/evaluation capabilities, error handling, and LLM output validation are all done very nicely with Pydantic AI that are often times missed or messy in other frameworks.
All of these features are more important once you are past the prototyping phase for your agents, so initially it might not feel super different than other frameworks. But once you try to really get your agent ready for production, this framework really gives you what you need! I hope that makes sense!
@@1brokkolibaum
To add to the "why".
Context Management: Maintains coherent interactions across multi-step tasks.
Example: A customer support chatbot remembers a user’s issue across multiple messages to provide a complete resolution without asking repetitive questions.
Logging/Monitoring: Tracks behavior for debugging, insights, and real-time issue detection.
Example: Logs show an e-commerce agent’s recommendations, allowing developers to identify and fix why irrelevant products were suggested.
Testing/Evaluation: Validates performance across scenarios and ensures changes don’t break functionality.
Example: Simulate scenarios where a finance bot calculates mortgage rates to ensure accurate responses even for edge cases like zero interest or maximum loan terms.
Error Handling: Prevents total failure with robust fallback mechanisms.
Example: A task automation agent encounters invalid input and gracefully alerts the user instead of crashing, offering guidance for correction.
LLM Output Validation: Ensures responses meet required standards, reducing unpredictability.
Example: A legal document drafting bot ensures the output follows proper formatting, avoids factual inaccuracies, and adheres to specified legal templates.
flowise with langfuse is incredible, just saying
It is! Though Flowise doesn't have all these production grade features that Pydantic AI does. It's great for prototyping though - I love using it for that and that's why I made a video on Flowise recently! I'm sure Langfuse helps with some of what is missing from Flowise, but I'm thinking it wouldn't quite be enough.
@@ColeMedin what are you getting from pydantic that flowise + n8n + langfuse doesnt offer, im curious its all relatively new to me
Phenomenal
Thank you!
Why show off with developing new code, when you modified their example? If you just modify their example code, It would be more useful, easier to remember, comprehend ...
UA-cam is free. Feel free to make your own video...
I actually very much took my own spin on this example - I only used their example as a base but I changed a lot and added the whole web research part of the agent, running with Ollama, and the Streamlit UI!
The best ai builder out there is lovable ai by far better than bolt I'm using it right now to build a spotify mixed with iTunes site what's your go to ai builder😊
Yeah I have heard great things about Lovable! These AI coding assistants are quite different from agent frameworks like Pydantic AI though, one is using AI to code, the other is a framework to help you create agents.
2 different things
Nice video. But the title is more clickbait than what the vidwo really can deliver
How? Are you slow? Its 100% on topic.
Wth is your problem?
Thank you! Honestly I stick to what I said in the title, I do think that Pydantic AI offers a way to build agents that surpasses all other frameworks. Not necessarily in getting started with building agents, but more making them production grade so you can actually build a product around them!
my only problem is using brave search api and not using duck duck go................ahhhhhhhhhhhhhh.why
You can easily replace it! I just like the functionality of the Brave API a ton