Crew AI with João Moura

0:00A-Gentic AI is seen as a key frontier in artificial intelligence, enabling systems to autonomously act, adapt in real-time, and solve complex, multi-step problems based on objectives and context. Unlike traditional rule-based or generative AI, which are limited to predefined or reactive tasks, Agentic AI processes vast information, models uncertainty, and makes context-sensitive decisions, mimicking human-like problem-solving. Crew AI is a platform to build and deploy automated workflows using any LLM and cloud platform. The company has rapidly become one of the most prominent in the field of Agentic AI.

0:38Joe Mora is the founder at Crew AI, and he joins the show with Sean Falconer to talk about his company. This episode is hosted by Sean Falconer. Check the show notes for more information on Sean's work and where to find him.

1:03Joe, welcome to the show. Hey there. Thank you for having me, Sean. Very excited to be here. Talk about AI agents. Yes, yes. Yeah, before we get there, so you're the CEO and founder of Crew AI. What's the story behind the company? How did all this start? I got to say, Sean, like people ask me this and my wife hates when I tell it just because it involves her. But I was working Clearbit for five years prior starting Queer AI. And in Clearbit, I was leadering their enterprise product and also all their AI initiatives. I was there for like all the way to the acquisition by HubSpot. And when we got acquired by HubSpot, I was already playing with agents in my personal life for myself, basically building agents to help me kind of like do better at LinkedIn and do better at X.

1:47My wife was telling me, you're doing so many cool things at Clearbit. You should be posting about it, right? But I'm very good at cheat posting on X, but I'm not that good on posting on LinkedIn. So I got a few agents to help me out. And I got hooked. It was so easy. I was getting so many posts in a very consistent manner. And I was like, you know what? I want to build more agents. So I looked around. I found nothing that could help me do it the way that I wanted to. So I decided to build my own. And that's kind of like how the open source started. And from that to make it into a company was just a very organic process of like companies reaching out to me and saying like, hey, we're actually using CREI in production.

2:27Can you help us out with this? And I realized there's no way for me to provide that level of support, like by just making it an open source project. Like it needed to have like funding behind it to make sure that we could do it. And that's how things came to be. And it got very excited after that. That's really, I think, the dream scenario for almost anybody who like ends up starting a company. It's like, you know, build something that's useful for you. Starts as like the side project that's doing, you know, something that you're happy with. And then suddenly people are like knocking on your door and saying, hey, like we want to pay you to use this project or we're already using this project.

3:03I mean, it's awesome. I think it's interesting, you know, what you said around doing some of these agents to help you do some of your work, like posting on LinkedIn. And actually for myself, like one of the first public AI based applications I built also was for helping me analyze podcasts that I do and then generate a first draft LinkedIn posts because I don't always have time to put those things together, but I want to do the guest right and help promote those things and so on. So that was like one of the first projects. Now, I did not, at least as yet, turn that into a business. So it's amazing that you're able to go from those humble beginnings to actually creating something of real value.

3:39Yeah, I gotta say for me back then was just like I want to make sure that I was finding the time to post this But like as you know, if you want to write something that is really good Like that has that like emotional connection that taps into your like experiences It takes a while like you can take a few hours for us to think like something true, right? So the thing was this would help me get from a crazy kind of like one two line idea to kind of like a two, three paragraph post that then I could basically like improve in five minutes and it was ready to go. And it was funny because I preloaded those agents with some knowledge.

4:15So they knew about my resume. They knew about my like life experiences. I wrote things about the day that I fell for my bike or the day that I broke my jaw and now that, and I got all those things in there. So it would draw all those interesting correlations that I think make this story very special that then I could spend like a few extra minutes making sure that just hitting the points that I really wanted to. But that was great, yeah. I still remember being kind of like on a meetup in the Bay Area. I think it was on Shack 15 and people from Oracle reaching out to me and say like, hey, so we're using Crew AI in production.

4:49And that was a big aha moment for me. I was like, what? Oracle is using Crew AI? So yeah, from that point on, things starts to really like basically get a lot of traction. and honestly I'm very grateful if anything as you said I know like how lucky I am from getting to work on something that I love so passionately and also like the fact that we have such a nice and incredible community behind us so I take none of that for granted yeah that's awesome yeah I feel like the company itself and the product has kind of jumped onto my radar maybe in the last six months and I feel like I'm hearing more and more has you know the growth been crazy during that time, are you seeing essentially this wave that's happening around agents and interesting crew?

5:32Yes. I got to say, it's like record after record. We keep dropping records every week or so. So it was funny because on the Christmas week, we broke our record in how many crews we executed in one single week. And each crew here, for the people that don't know about this, is a group of agents, right? So you could have two, three, five, seven. I have seen crews with as many as 21 agents on it. In this crew, the reason why these agents are grouped together is because they're trying to fix a specific problem or they're trying to automate a specific process. So that is why they're packed into a crew.

6:11So on that week of Christmas, we ran over 3.5 million crews. so those were kind of like tens of millions of agents in one single week and that was an all-hackered high but then fast forward a few days and I think a couple days ago in one single day we ran 1.3 million crews so like it's insane how much traction things are getting and it's funny because there's the open source there's an enterprise there's socials so there's all those different components that plays into the company. But I think open source and the enterprise adoption has been very impressive. And I think these years, it's going to be a little insane if things keep moving the way that they are.

6:55Yeah, that's amazing. So I think that's a good point to maybe just slow down a little bit on where we're going, just for people who are a little less familiar with this topic and talk a little bit about, you know, what is an agent and maybe get your perspective on this. Because clearly agents are having a huge moment right now, Even if you're not in AI, I'm sure you're hearing all about it. And people are making grandiose predictions for 2025 about that. But there's a variety of definitions out there. You can get everything. I've seen definitions of agent that basically suddenly everything is an agent.

7:25There's nothing that's not an agent. And then there's really specific definitions like agents have to be using tools. But in AI itself, there's a long history of AI agents. It's not just about gen AI. Like there's, you know, even going back 1960s and 70s, you had rule-based systems. It wasn't general AI or LLMs, but you had this notion of an agent doing, you know, some work for you. So how is, like, is the definition of a Gen AI agent different from the historical definition or is this just, you know, a different algorithmic approach that's happening behind the scenes to power the agent to capabilities?

7:59And how do you think about the definition of an agent? Yeah, no, Sean, that's a great question. I think it's something else. I think it's something different. When you think about AI agents nowadays, I think it's mandatory for you to have the AI dictate the flow of the application. So if you have a traditional software and you basically have a class with functions and the functions are calling each other and you have kind of like an API call that calls AI in there, that for me is a workflow. That's a script. That is not an AI agent. Now, if the AI, the LLM that you're powering this is actually controlling the flow of the application, and by that it has agency, then for me, that is an AI agent on its current definition.

8:48So yes, I mean, there's workflows, there's automations. Like, honestly, if you forget the name AI agents for a second, we're talking about AI powered automations. That's the book of it, right? But I think what throws an extra spin on it and make it interesting for unlocking automations that were just not possible before is this ability to not only generate the content in real time, but also change the flow of the application in real time. So it's not always a predefined path of graphs and nodes where it's doing always the same thing, but you're extracting the full value of the LLMs by having the LLMs itself control it.

9:25Right. Yeah. I mean, historically, when we think about like engineering anything, it's essentially a person that is coming up with like, you know, if this do this and this and this, and you're creating sort of this deterministic flow. But with the agent, you're sort of handing off that decision making process to the agent to figure out what is the plan of execution in order to accomplish a goal. Like I'm giving it a goal. It's just like you would give a person a goal and they go and they essentially figure that out. Yeah. And I would say that it's more of a spectrum more than anything else, what we're seeing out there in the real world for the thousands of use cases that we have been building so far is that it's not like one or the other.

10:04It's a spectrum depending on how much precision you need for your use case. So if your use case require less precision and you can rely more on those models, usually that means that you're generating more value just because you can have like less people overseeing this, these things can run over and over and it's on. And then you can have that, like, just go for it. But sometimes for certain aspects of the automation, you want to have more control. You want to have programmatic control of like what it's going to happen. So we allow you to do that as well on the feature that we call Create Flows, where you can basically say, all right, if this, then something else is going to happen.

10:42And that's basically a producer and kind of observer event listening approach. And that makes it extra flexible because you can decide when you want to have agency and when you want to have like more control. So I would say it's more of a spectrum than one versus the other, but definitely depending on the use cases and the precision that you're trying to get, you might optimize for a certain aspect. Why do you need something like a framework like Crew AI to build an agent versus doing something, you know, bare bones yourself? Yeah, I got to say, I love this idea on engineering that you should try to do the least amount that you can get away with, right?

11:20That's always a good way for you to get started with things. Don't over-engineer things. I always think from the ground up, what is the minimum that I can get away with it? And that's usually a good baseline. The problem is what we're seeing out there is these agents start very simple conceptually because they're like, all right, it's an LLM in a loop and I'm going to give it a prompt and it's going to do something. And then you start to think about the tools. You're like, all right, so I need to give it a tools. That's not a big lift. I can do tool calling. I can do JSON parsing. All right, I got that done.

11:49Now you're going to run this at scale. And they're like, well, if I'm going to be using tools, I need to have a caching layer. If I have a caching layer, I need to have an expiration mechanism for this. And if I have two agents and they're working together, I want this caching layer to be shared. Well, if there are two agents, they probably need to share a memory as well. Well, if they're doing memory, maybe this memory should be a rag so it can select the and then things start to get more and more complex. So I think where this framework's come in handy and I think where CREI really shines is it abstracts away a lot of that and give you kind of like a DSL that allows you for more simpler use cases.

12:27You can just use the DSL, but if you want to go super low level and change the prompts, changing the templates, changing all the inner workers, you can do that as well. but it checks all these boxes so that you have a gen delegation ability to communication caching tools and memory all that for granted so you don't need to rebuild those same things over and over and over and the other thing as well is as patterns are forming like you might try to do something for your company but the market is moving so fast that you're kind of like doing something that is suboptimal and a framework that's getting exposed to like a thousand companies are going to be able to learn from all those use cases.

13:09So those would be a few things that really can make a difference for some of these enterprises when they're using crew AI. What about in terms of the challenges around getting the abstraction right? All this stuff is so new, things are moving quickly. So you're going to have to make a decision. You're sort of having an opinion about how people should develop agents. And then you're encoding that essentially into this framework. and you're saying, this is what our opinion is about how to build these things. But how do you know that your opinion is the correct one? That's the thing. Early days, you don't, right?

13:43And that's the beauty of it, if you ask me. Early days, you don't. But then as people start to use it, you start to basically polish these rough edges and the ideas, you start to have a more clear take on what works and what don't. I think one thing that we did with Crew from the early days is a very opinionated way. And I think even how fast the market is moving, that's what people need. People want to have an opinionated way of like, how should I be building this? Like, what is actually working out there? Like, I don't want to try to think this from the scratch. So I think the fact that we had an opinionated take, if anything, helped the framework to get as much adoption as it did.

14:20This idea that we from the get-go are saying like, hey, this is what works. I remember looking back, a lot of the inspirations technically from the DSL came from the Ruby community, specifically Ruby and Rails community. I spent quite a few years doing Rails in my career, and I always loved how they prioritize developer experience and how there's a big piece of the language and the framework. And some people might say that sometimes it's too magical, especially Rails, and I agree with that. But I think there's something pretty about the ability to make something that almost read as plain English to the point that democratizes accesses to that technology, right?

15:00So the thing that we're trying to do is as long as we give you ways for you to go as low level as you want, is okay to have something that is more high level and you can basically figure out pretty quickly. Yeah, I do think that especially with framework adoption, like, it depends a little bit on what it is and what problem you're solving. But a lot of times, DX is a big factor in terms of like something catching fire and people wanting to adopt it. Because it's like, especially when we talk about something like agents, which is probably not necessarily something that most people have spent decades working in.

15:33They really just want to be able to solve their problem and get going and get to that aha moment as quickly as possible. So if you can lower the barrier to entry and the amount of friction involved with getting them to that hello world moment with an agent, then of course, they're going to want to do more and invest time and get into more complicated use cases. But you need to solve that problem to start with. Yeah. And I guess the other thing that adds to that is just you've got to use it. You can't build something and not use it. So I think from the get go, we have been using crews so much within our company ourselves that it's just like, it gives you clarity, right?

16:10Because you know what works and know what doesn't. You see like when you're struggling, you see new engineers during your team and non-technical people join your team and trying to build crews and update crews. So it gives you the clarity that you get when like you're a major user of your own product that I think really, really helps. And I think honestly, the community is kind of like, it's a flywheel. Like the more people you get in the community, the more feedback you get, the more you see what works, the more you can like policy the edges. And I think the other thing that also plays a role in this is because I have experience with enterprise software and a lot of our team also does, I think that we understand that you can't just move fast and break things on enterprise, right?

16:54You can absolutely move fast, but you got to not break that many things. So I think being able to make things backwards compatible and really fighting a lot for not having breaking changes, I think those are things that really also help drive adoption. Just because people know that the docs changed, they know that the features are coming, but there were very few times where we actually shipped something that was like, this is a breaking change. You're going to need to update your code to make this work. Okay. Can you break down what is the anatomy of an agent typically with Coit? Oh, yeah. So usually, again, there's an LLM in the middle, serves kind of like of the brain of the thing.

17:33There is a task that you're giving for this LLM. This LLM in Free AI, it's impersonating someone or a role. They're behaving as a certain role and trying to try to accomplish this task. And there's going to be an output. So that would be the core. Then you have tools. And you can think about tools as integration. Like it can be API calls, database connections, RAC implementations, whatever it might be, SAP, Salesforce, SharePoint, you name it. Those are integrations that your agents are going to be able to call on their own as they basically try to accomplish that task. Then if you put those agents together into a crew, they now automatically have the ability to use certain special tools.

18:18Those tools are delegating works for one another, asking questions to one another. And then because they are using all those tools, they get that caching layer that we were mentioning earlier. So now if they use the same tool, passing the same arguments, they're going to hit this caching layer. So now because they're all in the same crew, you're going to have a few different types of memory. You're going to have like a shorter memory. And this applies to any frameworks out there. A lot of these concepts are cross frameworks, by the way. You're going to have a short-term memory that basically allows your agents to remember everything that everyone is doing during an execution.

18:54You're going to have a long-term memory that is basically going to be improving your agents across many executions. So as they learn about what they did wrong, what they did right, they basically kind of like improve over time. And then we also have an entity memory so that they remember specific definitions of things. So for example, if they learn what is a software engineer, they won't need to have to try to learn that again. They basically, they already know about it. So those would be kind of like the memory. The only other thing that I would say that adds to this is they're usually, when you're bringing this into the real world applications, there's usually a trigger and a destination, right?

19:34Those things don't run on isolation. So there's usually a component of like, oh, this is gonna be kicked off when a new entry in my HubSpot appears or a new Zendesk card is created. And the destination is usually like something else or the same, like, oh, it goes back into my Zendesk or it goes back into my Salesforce or it goes into my email. So I would say high level, those would be the major components that you see in there. But then again, things can get very complex as well if you just keep working on that. So, you know, one of the things you talked about there was how an agent might have a role.

20:11And a lot of times you're configuring an agent role through things like the system prompt, where you're saying, you know, you're an expert researcher, expert copywriter or whatever it is. So, you know, what's happening underneath the covers with the LLM when you give direction like that? Like, is that directing the LLM to a certain set of parameters or adjusting different vectorized terms that influence each other to change how the LLM is going to like interpret and respond to things? Yes, a thousand percent. I think a lot of people that are using these AI APIs right now, they have a high level understanding of how these models work behind the scenes.

20:45But honestly, if you look back to any AI model out there or most of the AI models out there before LLMs even, if you think about prediction models or classification models, in the end of the day, they're trying to all do the same thing. They have something that they want to predict and they have other what people call features that they already know. That's what people call in data science or ML. So you have the features that you know when you have the one thing that you're trying to predict. So a lame example would be, what is the likelihood that's going to rain? You don't know if it's going to rain or not, but you know what is decision of the year, you know what is the temperature, and you know your location.

21:22So giving those, if you have enough data sets from three or four years, you could pass this into a model that is going to try to create a mathematical function that giving the data that it knows, it can predicts the data that is unknown, and that's the likelihood of writing. So that is any model out there. It's what they're trying to do. With LLMs, they're just not different. The only thing that changes is that in this case, the data that you know, the features, are the text that you have written so far or the tokens that are outputted so far. And it's going to use those tokens to try to predict What should be the next token?

22:02So the more information that you put before the token that you're trying to predict, or the more qualified information, the better steering power you will have on the next token. So if you say it's 37 degrees Celsius on the winter, but if you change it to a summer, that have a directly impact in the likelihood of rain, right? So I would recommend many people to try this out themselves they can go into chat gpt or claude or whatever it be and ask for like give me a stock analysis on tesla and then open a new chat and try the same thing but it starts with you're a finna approved investor and you're going to see they're going to have wildly different responses because there's two in the model in a different way so role playing can definitely play like a big role on how those models perform if you do it right Right.

22:54And in terms of the long-term memory versus short-term memory, how that works behind the scenes is essentially you're automatically populating the context. So you're doing some context data augmentation before whatever the new behavior is or the new event sequence that the agent was execute. But for the long-term memory, you're just holding on to that as persistent. So it's always there within the context window. Is that the idea behind the long-term memory? A little bit. So the long-term memory, it's actually a Grag data because it uses Chroma. behind the scenes, but you can customize if you want to.

23:27And the reason why we can do something like that is on Cree AI from the day zero, we made sure that when you're defining a task, and it was funny because when we created this back then, a lot of engineers didn't understood and kind of gave us trouble because of it. When you're creating a task for an agent, you have to not only describe the task, but you have to say what is the expected output. A lot of engineers got mad at us because of that. They're like, well, I'm already given a task description. Why do I need to say what is the expected output? What they didn't realize is because, one, we're forcing you to do better prompting, even if you can tell.

24:06So if you haven't been doing much prompting up to this point, this is kind of like a forcing function for you to think like you're supposed to think in writing the prompt. But then we can programmatically use an LLM as a judge to compare your actual output with what you said the expected output would be. And then we can extract learnings from the discrepancies in there. And then we can save that into a vector database. And then your agents can query those learnings during execution to learn from what they have done wrong in the past so they can correct. So a lot of the long-term memory is memory that doesn't apply for that specific run that you're doing of your agents with that specific data that you're putting in through that.

24:50But it's more about the way they're supposed to behave and things that they haven't complied with in the past that you expect them to. So they start complying with that over time. So the long-term memory is injecting extra context with learnings from discrepancies between what you expect them to give in the past and what they actually gave you. but everything happens automatically. In terms of like learning, are people actually training agents in the same way that you train like a, you know, fine tune and LLM or are you primarily relying on this prompt augmentation to provide the right context?