Sign in to continue learning
Design, Develop, and Deploy Multi-Agent Systems with CrewAI
Beginner
3 hours 45 mins
Topics
Agents
Now, I want to make sure that we talk with people that are building AI agents in the real world. I want to talk about how these people are using CrewAI to build these agents, what are the things they're finding exciting about, what are the things they're having challenges with, and everything in between. So in this lesson, we're going to chat with these people, they're leading these projects and making huge strides on the adoption of AI agents in their companies, and actually getting amazing results. So let's make sure that we dive into this so we get to know more about how they're expanding their usage, how they're finding new opportunities, and what are some of the lessons that they learn that you might be able to apply to your own use cases. So let's join that conversation. Now you're going to hear from Exa. Exa is a major partner for CrewAI. Exa is going into the search space, building their own searching engine, optimized for AI and AI agents, making it easier for people building use cases to get the right data on the right time, exactly how they need it. Exa integrates with CrewAI on an official integration that we have built and you have been using through the entirety of this course. It processes millions of requests a month for AI developers, many of those powered by CrewAI agents. And enables agents to access real-time information through semantic search. So let's hear from Jeffrey, one of their co-founders, on how they're thinking about AI agents and specifically how they're feeling about their role that CrewAI is playing into the space. Jeff, thank you so much for sparing the time to be with us today. I know that it's very exciting to talk about AI agents. There's so much that is happening right now. So as the agenda in space continues to evolve, how important it is from your perspective and what you're seeing, this broad ecosystem of tools and resources to keep expanding, let's say. I think that it's been really incredible what's happened in the last year. Basically, at one point in the past, the LLMs were only, at one point, they were really, really bad at using tools. And then they got kind of okay at using tools. And now they're really, really good at using tools, right? Like Claude 4.5 sonnet and GPT-5 Codex, like they're just amazing. The LLMs, if they don't have any interesting data put into them, or if they cannot engage with the external world, it's like if you're a human and you're in a room and you don't have a computer, you don't have anything available to you, you're just kind of sitting there. And you're smart, you have a brain, but you're just sitting there. And so fundamentally, it's critical for LLMs to be able to engage with the external world, read information, and also write information. At least from my understanding of what the AI labs that are working on and what kinds of capabilities they want to be enabling in the coming years, it's largely focused on tool use, external data stuff. You know, a thousand percent, and it's fun that you put this. I was actually talking with Andrew about this as we were recording this course, and he mentioned that like you go back 2024, what feels like a decade ago in AI times, and these models, like they really, they were not that good. Like there was so much scaffolding that you had to put just to make sure that they would behave in a certain way. You had to describe every single step to have all the control. And now that models are getting better and better and using these tools and all that, it's really causing people to remove a lot of that scaffolding and doubling down a little more on the agency side of things. So how do you see specifically with Exa, because you have this very interesting perspective on real-time web search, and that unlocks so many different use cases, right? These agents can do so much if they type in the real data and write data at the right time. So how do you see these autonomous agents now evolving, given that they're getting access and being paired to these real-time web search APIs? So when GPT 3.5, if you gave it a few thousand tokens and told it, hey, answer this question from this webpage, like 30% of the time it would hallucinate. Like it would just tell you like the wrong answer from the webpage, right? And then at a certain point, maybe this is like GPT-4, it started to hallucinate a lot less. But still, if you gave it like multiple options, like, you know, hey, give it like two tools to use or something, it's still not going to perform well. And then now, you can build complex agents with access to say, you know, a dozen tools and combine the use of the tools in all sorts of crazy ways. For instance, that's what Claude Code does, right? Like Claude Code has like 10000 tokens worth of tools. It's the file grab tool and it has a web search tool, it has all these things, and it's able to chain these tools together and produce really useful outputs. Web search is just a fundamentally important part of that tool chain because the large language models fundamentally, they're not search engines, right? They know a lot of stuff, but it's just because it's just the stuff they've memorized and sometimes they hallucinate. So they depend on tools, including web search. Yeah. And I think you have a good point because it's not only the ability to web search, right? Your point on hallucination is almost like grounding its answers, right? And the fact that you now can give this ability for these agents to like not only give you an answer, but fact check themselves or make sure they're working with the latest data and they're finding everything that they need to check all the boxes. I think that's something interesting. Is there any specific and like use cases that you have seen around this idea of like, hey, this is where grounding in real data and updated from real time web data really makes a huge difference. And what are those use cases that are seeing out there? Okay. So let's just think about like what LLMs haven't memorized, right? So like one is recent things, right? So so if you're if you're asking LLM, like what happened today in the world, it's not going to anything unless it looks up information. The second is long tail information. So like, for instance, say you're like an AI to go market tool, you're building like a sales prospecting tool or something like the LLMs just haven't memorized like a list of 1000 biotech companies or like 1000 AI startups or something, right? This is just information it doesn't know. And so, you know, there's a long tail of information that's not super common that everybody just knows. Like, you know, Stripe, right? But it doesn't know, like the latest YC company that was working on payments. And so it needs to look things up in these cases. And then finally, I think one really interesting thing that we've been focusing on Exa a lot is actually coding, like a huge limitation of the LMS right now is that they can code really, really well for things they've memorized, right? Like next JS, they can write perfect next JS, they can write perfect express, like whatever. But, you know, say that like last week, OpenAI released, you know, agents SDK, and then also a few other SDKs. Well, these LLMs, even OpenAI's own LLMs don't know anything about these SDKs. And so we fundamentally need to do web search to find out information about how to code. And so we actually launched something called ExaCode a few weeks ago, which is, for instance, like a coding agent specific search. I love that. And it's funny that you mentioned that. That's one of the major use cases we have with PwC, one of our customers, where they want this model to basically do ABAP and APEX, so SAP and Salesforce coding. And this model, I mean, there is no sum about it, but not about the PwC way of doing so, right? Or the conventions that they have viewed across many build outs from this application. So yeah, that's a great use case. I mean, coding is not the first thing that would come to mind for most people when thinking about this. But you're right. Like if you're not using next JS, like what probably these LLMs know a lot about, you need to tap into like real grounded data. That's great. From a technical perspective, like what do you think is the biggest friction points and like people are really trying to like adding these embedding searches or doing crawling or whatever it needs to get into that data. Like do you think that the tooling side of it spit you it out? Or do you think like it's like the quality of the outputs or like what do you think is the main kind of like thing that is getting in the way nowadays? I think the one correct answer to what is limiting the agents from using like search tools and crawling tools and whatever is context window management, like context engineering. And so like today, if you are using a lot of these like tools that are like sort of popular, what happens is like, for instance, today, if Claude Code uses this native fetch tool instead of using Exa MCP for web search, it downloads like files that are like hundreds of kilobytes and just completely floods the context window. And I think that people kind of have maybe forgotten how fragile the context windows are because the models feel smaller. And then it's like when the model stops working, you're kind of like, oh, OK, it didn't work. Maybe I'll try again. Right. But like really, the models just start not performing well with the latest models and the most powerful models, even at like the tens of thousands of tokens level. And so if you flood a context window with like 10000 tokens of like raw HTML or something like it's not going to perform well. And so I think it's really critical for these tools to kind of do some of the token efficiency steps on their own. And so, for example, ExaCode has a philosophy where we always try to return you just a few hundred tokens. We actually try to just return code examples because code examples are the most dense representation token wise of what you need to know to do your coding task. And if not, then we pull documentation. But even then we try to keep it to low thousands of tokens. And this goes for any tool that you're creating. Like if you're like ingesting like gigantic documents somewhere, it's going to flood the context window. So maybe you have to use sub-agents or something, right? Like, for instance, Claude Code has sub-agents now, which is like super cool. That's a good point. I think like it's funny you go back into like Google when Google was the de facto way that people are searching things. Right. And there was this metric of like, oh, we want to be the website where people spend the last time on. We're like, well, all the other websites building to like optimize the engagement time. They just want to be that quick bridge that you find everything that you want super fast. It feels like what you're telling me is that on the new age where like Exa now is being like a huge provider for these searches for agents or new version of that is I will return the most amount of information, the last amount of tokens as I can. So like really getting the most bang for your buck, not only in terms of money, but like in the context window, right? Because in the end of the day, the context engineering is actually something that you really got to consider. Like these things are still AI models at the end of the day. It's like garbage in, garbage out. So the ability for you to do effective context management can have a good impact. Well, I know that we're closing at time. I do have one final question. Any takes on kind of like what we have seen with CrewAI and Exa out there, any specific use cases that you have heard about or things that you have found interesting or any takes on kind of like the AI agent space as abroad and kind of like how things are going? Products and frameworks that are maximally flexible, given the increasing capabilities of the large language models are increasingly important. So I think that like where I've seen really cool use cases involving CrewAI and Exa is like, I think both CrewAI and Exa have this philosophy. It's like, hey, like you have a really smart LLM, right? Give as much power to the LLM to be intelligent and do things. Try to leave it alone, right? Like build, build nice things for it to use, but don't like interfere and like try to guide, like try to control it in this very specific way. And so all the coolest things I see built on Exa is just like, hey, like, you know, you could use some sort of product out there that kind of has some very like narrowly defined use case or something. And maybe you're trying to build something for yourself that's like that, but like give power to the LLM and like let it breathe and like use things flexibly. Yeah. I love that. Give power to the LLM. I'm going to use more of that. Jeff, thank you so much. I really appreciate you spending time with us today. This was extremely helpful and I'm sure that people are super excited about everything that they use Exa for in this course. I'm sure they're going to be using it even more after this. For sure. Yeah. Thanks so much for having me, Joe. It was nice.
