What are AI frameworks? And do you need to care which one?

Hey, my name is Mahan, the CEO of Nara. My name is Luka and I'm the CEO of Nairon. Luka, over the past couple videos, we've gone through the AI fundamentals, slowly gone into more agentic concepts, and in today's video, we're going to cover AI framework. So, for the audience to get a easy introduction into what AI frameworks are, let's start with an analogy and take it from there.

For sure. So an analogy that we like to give is that every good restaurant has a really good chef but also a fully stocked kitchen. In a kitchen you have your refrigerators, your ovens um and all the other tools that come along with a kitchen. In this case when we are building an AI agent which is essentially like a dish in this analogy um it's obviously important to have a really good chef which is the model the intelligence behind how to actually produce a dish in this case. Uh, but you have to have like a really good oven and you you have to have a really good refrigerator to be able to execute the work. And so instead of building your own refrigerator and building your own oven, you essentially buy a refrigerator and an oven that works. And you don't really bother yourselves with the intricacies of how exactly an oven works or how exactly a refrigerator works. Uh, we'll get into more details exactly, but I want to give you a very quick idea of some of the biggest players in the space. Um the first sort of AI framework that came up uh was lang chain really good up until today. Um there is a learning curve as a developer on how to learn lang chain and then how to build AI agents but it's a very good generalpurpose agent builder. You could say lang chain has a bunch of little assets that you can pull from um if you need that certain type of functionality for your agent. Other ones as well as by the U Lama family is the Lama index.

That's really good if you're looking to do anything that is search related or if your agent has a lot of uh search related tasks that it needs to do. Um really good to use some of their assets. And then there's others lesserk known ones um like Crew AI and Autogen and some of the really big model companies like OpenAI and Anthropic. they have their own AI frameworks that they've used to actually build their own agents like cloud code uh which you can essentially use them. If we look here a little bit more detail on what exactly those like pre-built tools are you can see and I mentioned already the ability to actually connect to a model or for example the ability to route to a different model for a different use case or perhaps when a model is down. So as we might know um sometimes Entropic and all these other bigger model companies sometimes their models experience performance degradation. The latency is increased um the model is not providing as good of an an answer. Um sometimes we need to quickly shift over to another model without the end user actually knowing that now we're operating on a different model. Another example is managing a conversation. So you could use an agent for five minutes continuously, step out for an hour. The agent is not essentially going to be running waiting for you every minute um and every second to reply. It'll essentially sort of like cool down waiting for you to come back. That sort of like engineering plumbing work is all done by these libraries. Another one is like handling tools. We spoke about MCPs, CLIs, and APIs. Um and all of those functionalities bring a lot of errors and error handling is one of the biggest things that is being provided by these uh by these frameworks.

Now that we're covering more complex topics, it's always important to take a step back and actually bring the topic back to how it affects day-to-day business owners. So why should a say non-technical operator even concern themselves with an AI framework?

Well, essentially they don't need to. It's not really a decision on the buy side of the process. um you're not buying a specific agent because they're built on a certain framework. You're not even using an agent because they're built on a certain framework. Most of the times you don't even know. Um however, if you're looking to build this in house, that's when you should concern yourself. It's really a developer question. So, what are my developers familiar with? What functionality do I need? And what are the limitations of some of these frameworks? So, say for example, my team is wellversed on Crew AI. Well, we should just use crew AI because it covers essentially 90% of the same frameworks. Now, another very important thing is that sometimes on very rare occasions and this is why planning is important, the frameworks can be very opinionated and that will actually shape how the agents are built.

So, if you need a very quick agent um with not a lot of overhead, you might choose something else rather than lang chain. However, on the other end, if you are building a multi-purpose um agent, then you would go with Langchain cuz it's probably one of the most uh profound frameworks out there, it's it's also the one that most engineers know and are very comfortable with. That's also another thing. You don't want to jump on the latest framework because many of the times the companies behind those frameworks are typically doing it for free or like a lot of stuff in in software open sourcing the frameworks.

And so you don't want to be behind a project that might get, you know, discontinued uh or depreciated. So that's essentially why a business operator would need to know about frameworks and why they would even concern themselves.

So that's just really been a brief coverage of what AI frameworks are. And as you mentioned, it's not really something an operator is going to concern themselves with on the front end, but rather something a developer is going to wrestle with on the back end.

Yeah, I mean it's really important to understand how much work goes into actually building these agents even at the most basic level even at the level of the agents knowing how to reply and how to use a model and all the use cases which we talked about. And so that's really just to show what the plumbing looks like. Yeah. And guys, if you enjoyed this video, found it helpful, you can subscribe to our newsletter. The links down in the description below. You can also find us on LinkedIn. We're very active on a day-to-day basis. And until next time, we'll see you in the next video.