AI Engineer World's Fair 2024
Using agents to build an agent company
Read the talk
Using agents to build an agent company
João Moura traces crewAI from a personal writing automation to company workflows, then shows how agent coordination, retained instructions, and deployment change what engineers must build.
From a talk by João Moura
Before you start: Basic familiarity with LLMs, APIs, and Python configuration is helpful; no prior crewAI experience is required.
Ten million executions, one simple agent loop
João Moura opens with a count: 10,602,922 agent executions through crewAI in the preceding 30 days, he reports. That is an execution total, not a count of unique agents or successfully completed business tasks. It sets up the engineering question behind the talk: what does it take to turn a model’s ability to generate text into automation that can keep working?
An LLM can generate content and choose among options. Moura’s simplified agent model adds a conversational loop: the model talks to itself or another copy, makes decisions, and uses tools without requiring the user to supply every next turn. The consequential change is that the user can leave the loop while the system continues acting. The opening diagram culminates in that transition from model interaction to an autonomous agent.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From connected paths to coordinated crews
A conventional automation often begins with a straightforward path from A to B. Add C and D, and the engineer must connect more paths and account for more circumstances. Agents move some of that decision-making into runtime: give them available options, and they choose according to the situation they encounter. The branching logic changes location; the engineering work does not disappear.
The initial anatomy—an LLM, tasks, and tools—quickly expands in production.
| Scope | Additional concerns |
|---|---|
| One agent | Caching, memory, training, guardrails |
| A crew of agents | Communication, shared cache, shared memory |
| Multiple crews | Coordination between crews |
Each level builds on the previous one. Connecting agents means deciding how their work and retained context fit together; connecting crews adds another level of coordination.
This also changes the software contract. In Moura’s conventional-software example, known input types feed predictable operations such as addition or multiplication, producing outputs that are straightforward to test. An AI application may still accept a string, but that string could contain a CSV, a recipe, or a joke. Its type does not tell you what it means. A model then processes that input through behavior the application cannot fully inspect, leaving the output less predictable. The distinction is semantic predictability, not simply whether the program has types.
Moura separately reports 100,000 crews executed daily. A crew coordinates agents, so this is a different unit from the opening agent-execution count. He presents crewAI as a library for building and orchestrating multi-agent automations and cites its execution volume as evidence of production readiness. He does not define a formal readiness standard or provide a success-rate measurement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A writing task becomes a reusable framework
The company began with a much smaller problem. Moura, crewAI’s founder and CEO, describes working at Clearbit when his wife suggested that he share more of his LLM work online. Writing LinkedIn posts was difficult for him, so he built agents to do it. The attention those posts received encouraged him to automate more tasks, and crewAI emerged as a way to reuse the approach rather than rebuild each new agent from scratch.
He developed the framework early in the mornings during an anniversary holiday. As others began using it, reports arrived of agents going down rabbit holes, hallucinating, and encountering tool errors. Community growth accompanied those failures: Moura reports more than 16,000 GitHub stars and more than 8,000 Discord members. A community member also created a Reddit community. Interest from engineers and companies turned a personal automation project into a company-scaling problem.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Drafts create attention; qualification prepares meetings
Moura’s first company workflow was a marketing crew with four roles: a content creator specialist, a social media analyst, a senior content writer, and a chief content officer. Rough ideas entered the crew. The agents checked X and LinkedIn, searched the internet for more context, and drew on Moura’s previous experience to produce drafts. Moura still performed the publishing step. The described workflow automated research and drafting rather than unattended posting.
Moura reports 10 times more views in 60 days. He supplies neither a baseline nor a controlled comparison, but the increased attention created a concrete next task: qualify the people arriving. He chose lead qualification as a higher-impact, lower-risk expansion beyond the initial marketing workflow.
The lead qualification crew combined a lead analyst expert, an industry researcher specialist, and a strategic planner. Its procedure was to:
- Analyze a lead’s responses.
- Compare those answers with CRM data.
- Research the lead’s industry.
- Produce a score, potential use cases, and talking points for a meeting.
The output was preparation for a human conversation. Moura could enter a meeting with context already assembled rather than perform that research himself.
Moura reports conducting more than 150 customer calls in two weeks. Qualification had moved work out of meeting preparation, but the calls themselves remained his responsibility. The next bottleneck was human capacity to serve the interest the earlier automations helped generate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Expand from a working task
The internal rollout continued into code documentation and email. Moura says agents write crewAI’s documentation; email is another named application, without a detailed workflow. He then presents company adoption, with a slide showing Carlyle, PwC, K2, KPMG, Aurecon, Capgemini, and Havas logos, and describes the companies shown as using Crew. He also cites Dharmesh, identified as HubSpot’s CTO, and Jack Altman as investors. These are adoption and endorsement claims, rather than descriptions of those companies’ implementations.
Moura predicts that agent adoption will not reverse and that its impact will exceed the internet’s. His practical advice is narrower: begin experimenting, start with something simple, then expand toward low-risk, high-impact work. The company’s sequence gives that advice substance: draft content first, prepare leads next, then extend automation to other recurring tasks. He turns from that operating pattern to the capabilities being released in crewAI.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let agents build tools
Agents need tools, so the next proposed capability is to let them write code themselves. Moura first describes an elaborate executor-instantiation setup as a joke, then reveals crewAI’s actual interface: the allow_code_execution flag. The slide places it alongside an agent’s role, goal, and backstory.
A minimal Python illustration of that historical configuration is:
python
from crewai import Agent
coder = Agent(
role="Python developer",
goal="Write code to solve the assigned task",
backstory="You build small Python tools for concrete tasks.",
allow_code_execution=True,
)
This is historical configuration: the versioned agent documentation now marks allow_code_execution and code_execution_mode deprecated and states that CodeInterpreterTool was removed.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Retain instructions and bring other agents into the crew
The next capability follows an employee-onboarding analogy: when someone joins a company, they receive training. Moura introduces a training CLI through which users provide instructions that become part of the agents’ memory for later work. He presents this as a route to more consistent results. Training here means retaining guidance, not updating model weights, and consistency is an intended benefit rather than a demonstrated guarantee.
He then announces support for bringing LlamaIndex, LangChain, and AutoGen agents into crewAI. The promised benefit is access to the surrounding crew facilities, including shared memory and common tools. This is a claim about incorporating third-party agents, not merely calling external tools. Moura says the new version shipped immediately before he came onstage and was available that day.
For a learning path, Moura identifies Andrew Ng as another investor and points to the crewAI course, which he describes as a two-hour introduction.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a local crew to an API—and a generated first crew
The closing problem is deployment: a crew running in a terminal still needs a way to serve other applications. Moura introduces CrewAI Plus, the enterprise offering at the time, as the path from local execution to a hosted service. The starting point is a working crew that the developer selects and pushes to GitHub.
Moura claims that a locally running crew pushed to GitHub can become an API through CrewAI Plus in three minutes. He describes that API as autoscaling, protected by bearer-token authentication, and running with a private VPC. The three-minute figure is a product claim for this deployment workflow, not an independently measured benchmark.
The next step is an interface: a one-click export to a React component supplies a UI for demonstrations and customization. The intended progression is therefore a local crew, a deployed API, and an application interface that can connect the crew to other software.
The launch offer promised CrewAI Plus access in less than 24 hours to the first 50 companies signing up through the presented link. That offer belonged to the announcement, rather than an ongoing access guarantee.
Moura ends by removing another setup step: building the first crew. He describes a crew that takes only an email address and company name, creates another crew, and pushes it to a GitHub repository. The generated crew can then become the company’s first CrewAI Plus deployment. Repository creation is the promised automated result; deployment remains the next step. The company-building loop closes with agents producing the starting artifact for another agent workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Official Python framework repository with installation guidance and examples for coordinating agents into crews.
CrewAI's learning page links to its multi-agent systems course at DeepLearning.AI.
Updates since the talk
Agent configuration reference covering tools, execution controls, memory, and code-execution deprecations.
Explains the training CLI, feedback artifacts, and how saved suggestions guide later agent runs.
Read the complete timestamped transcript
- 0:00
[upbeat music] 10,602,922.
- 0:20
Over 10 million and a half. That's how many agents got executed with crewAI in the last 30 days. [audience cheering]
- 0:31
And this is crazy. I'm so impressed by this because this is real. It's real, and it's moving way faster than most of people think. And I assume most of you here know what AI agents are, but I'm gonna catch you up if you don't real quick.
- 0:46
So LLMs, we all know them, ChatGPT. Turns out great to create content, and they're so good that they almost look they are reasonable. They can choose between left and right, and up and down, and all the other options.
- 0:58
And if you get them to chat with themselves or to a copy of them, guess what? You can leave the room. You have an agent. Basically, it can take its own decisions, it can use the tools, and it can be autonomous.
- 1:10
So if you didn't know what agent was, there you go. Now, you might be asking, "Great. Now what? What's up to me?" Well, we have been building automations as engineers for decades, and usually it starts pretty straightforward.
- 1:24
It's like, "Hey, I wanna go from A to B," but then what happens when you add C and then you add D, and these things can get complex pretty quick, and that's how legacies and headaches are born.
- 1:36
But, uh, turns out that with agents, you don't necessarily need to do that. You don't need to connect the dots. You give it the options, and the agents can adapt to the circumstances that they are met, and they can do that in real time.
- 1:50
So that allows them to build automations that were never possible before, that you couldn't do it. And when you think about the anatomy of these agents and what they look like, they might look pretty simple at first.
- 2:01
You might say, like, well, you have an LLM in the center and you have tasks and you have tools. But once you start to building these things in production for real, you quickly realize that you gotta think about, well, I need a caching layer, I need a memory layer, I need to train them, I need to find
- 2:16
a way to add guardrails, and so much more that goes into that. And then now you wanna like them to talk to each other, and that adds another complexity layer.
- 2:25
And then when they're in a crew, you wanna still think about the caching, but now it's shared in the memory and now it's shared. There's so much that goes into this.
- 2:34
And then you can go one extra level and get multiple crews to talk to each other. All that goes to say that the way that we have been building software is changing a lot.
- 2:43
If you think about the way that we used to do, it's very strong type. All the software that we have built is very strong type. You start with knowing exactly the inputs that are coming in.
- 2:52
It's a form, it's an integer, it's a string. You know what's happening. You're summing it up, you're multiplying, and then you have a very strong output to the point that you can write basically any tasks because the behavior is always the same.
- 3:07
But with AI agents and any AI apps, for what it's worth, everything is fuzzy. Uh, you don't know what's coming in. Yes, it's a string, but it can be a CSV, it can be a recipe, it can be a random joke.
- 3:19
And then these models are basically black box, and you don't necessarily know what's coming out of it. And you know what? I love it. So this is happening now, and I'm being serious.
- 3:32
Every single day, 100,000 crews are executed, and I'm talking like every day.
- 3:42
And I, I mean, I have been talking a lot about crewAI, and, uh, crewAI is a, a production-ready library to build and orchestrate multi-AI agents automations. And we'll talk more about that in a second.
- 3:54
And, um, uh, I don't know exactly what like, uh, is the, like the book definition of a production-ready framework, but I'm pretty sure that involves running more than 10 million agents every month.
- 4:07
So, uh, I, I like to claim that. And the fact that we have been building this means that we get a lot of exposure to a lot of use cases.
- 4:15
What are people building out there? How they are using this? And it's so good. Before I move along, really show of hands real quick, who has tried crewAI? Raise your hands.
- 4:26
I like this, but we are gonna get a even higher number. So I'm João. My name is kinda hard to pronounce. I go by Joe sometimes. Nice to meet you.
- 4:36
I'm the CEO and founder of crewAI, and the way that I built this company has been a very interesting journey. Everything start back in Brazil. I'm a long way from home, and everything started with my wife.
- 4:49
I'm very blessed to have a very smart wife, and I have been working Clearbit for many years before starting crewAI. And my wife told me, "You know what? You're building a lot of interesting stuff with like LLMs.
- 4:59
You should be sharing more about it online." But I suck at writing LinkedIn posts. So as a good engineer, I was like, "Hey, I'm gonna write some agents to do that for me."
- 5:08
It turns out it fucking work. [audience laughing] I got so many views. Everyone's going into my LinkedIn, and I was sold. I was like, "You know what? I want to automate my life away.
- 5:19
I don't wanna do anything anymore. I'm gonna just relax." Well, to my surprise, it didn't turn out that way. Uh, and I start to building crewAI because I want to use the same thing to build more and more agents.
- 5:31
The problem was all that happened in my anniversary.
- 5:36
So you can see that I was having a lot of fun, but at the same point, my wife doesn't like me to spend too much time in the computer when we're in the holidays.
- 5:44
Uh, but little did she know that I was working super early, and I was hacking away, and I was building crewAI. And things start to cook off, and we start to getting bugs, like rabbit hole reports, hallucinations, two errors.
- 5:57
I'm not gonna lie, things got a little crazy for a hot second there for us, but turns out with that also came a great community. Over 16,000 stars in GitHub.
- 6:07
Uh, a g- a Discord community with over 8,000 people. And then [REDACTED:username] create a Reddit. I didn't know, but we have a Reddit, and that's amazing. And then not too long after that, engineers start to, like, reach out, and then companies start to reach out, and I'm talking about hundreds of thousands of people, and I'm like, "All
- 6:25
right, we need to scale this up." But then how we do it? How do we scale a company in such a competitive market? Guess what? The answer was in front of me all along.
- 6:35
We need agents. [laughs] And so did I build some damn agents. So let me tell you how I start. I was like, "Hey, let's start simple." This is a company.
- 6:44
What do we need? We need marketing. So I build a marketing crew. I was like, "I'm gonna build first a content creator specialist, a social media analyst, a senior content writer, and a chief content officer."
- 6:58
Bring them all together. This is my marketing crew. I'm gonna shovel in rough ideas that kinda suck, and I wanna get something great. So they're gonna check X and check LinkedIn and what other people are talking about this.
- 7:10
They're gonna search Internet and learn more about the topic. They're gonna look at my previous experience, and they're gonna give me incredible draft, and then I start to posting it out.
- 7:19
And again, guess what? It worked. We got 10X more views in 60 days. 60 freaking days, and I was loving it. But
- 7:30
with that came a problem. I was like, "Well, I need to move on. I, I need to, like, now serve these people. How do I qualify them?" Well, I need to go the next level.
- 7:39
I did a simple crew. I'm gonna step up the ladder. I'm gonna do the next step, the higher impact but lower risk. So this is what I did. I did a lead qualification crew.
- 7:49
So I was right. I'm gonna bring up a lead s- analyst expert. I'm gonna bring an industry researcher specialist and a strategic planner. I'm gonna wrap them together into a lead qualification crew.
- 7:59
I'm gonna shovel in my lead responses, and I want them to analyze the answers. I want them to compare them with my CRM data. I want them to research the industry and give me, like, a score, use cases, talking points, so I can jump in a meeting right away.
- 8:13
Guess what? It working. The problem is it working too well. [laughs]
- 8:20
I end up doing 150-plus customer calls in two weeks. It was crazy. You know what? I don't regret it. I love it. So this is what happened. I start to expand more.
- 8:32
Let's build more crews. We have marketing. We have lead qualification. Let's do code documentation. So if you try Crew, all those docs, we didn't write it. Agents do it for us.
- 8:41
And I was like, "I wanna do more," and start to do email and do more and more, and it works, so to the point that these are some of the companies.
- 8:49
They're now building with Crew. They're using Crew, and it's insane to me. And hey, if you don't like these companies, you don't believe that, well, believe some of our investors, Dharmesh, CTO of HubSpot, or Jack Altman.
- 9:01
I mean, they can vouch for us. We are doing pretty well. And what about the future? Like, where, where is this thing going? Well,
- 9:09
as an LLM model, I can't... No, I'm kidding. [laughs] [laughs] So actually, the genie's not getting back in the bottle. This is gonna be huge, bigger than the Internet. We all know it.
- 9:20
Like, this, this is not going back. People are not gonna stop using agents from one day to the other. Well,
- 9:27
this is my advice for you. Be an early adopter. Don't wait for other use cases. Start simple. Expand to low risk and high impact. That's what we did. But hey, I'm not gonna finish here.
- 9:39
In CrewAI, we are known because we ship damn fast. So I'm gonna make some announcements of some of the stuff that we are working that I'm super excited about.
- 9:48
First thing, your agents need tools, right? So why don't you let them build their own tools? So we are working with code execution, what it means that in the new version, all you gotta do is create an instance of automated coder command line code executor.
- 10:04
Come on, you're not buying this. I'm not AutoGen or whatever other framework you're using. We're CrewAI. All you gotta do, it's one flag, allow code execution. That works. Our agents can code now.
- 10:13
You don't have to worry about this. [laughs] Another thing that we are working on, you know how you do when you hire a new employee? You train them. Why not do that with your crew so you can get consistent results over time?
- 10:25
Well, there's a new feature, train your crew. It's a new CLI. You can run that, and you can give instructions to it, and that's gonna become baked into the memory of your agents to the point they are gonna give you consistent results every time moving forward.
- 10:38
But we're not stopping that as well. There's more. You know what? I like to think about us as we don't see agent callers. We want all the agents, so bring them all.
- 10:50
We're a universal platform. Bring any third-party agent. You know your LlamaIndex agent, your LangChain agent, your AutoGen agent? I mean, I don't know why you would use anything else because you got Crew, but hey, come on.
- 11:01
You can bring them into the party, and they're gonna have all the CrewAI agents features, the shared memory, the same tools. You're gonna be able to use all of them if you want.
- 11:10
And then, again, the best thing about all this, you can try it today. We just ship the version before I come on the stage. And if you wanna try it right before this call or later in the day, you can give it a try.
- 11:22
It's a new version. It's live. And if you, again, that is not exciting enough, maybe you wanna hear from another of our investors, Andrej Ying. So you go and check it out.
- 11:31
We put together a two-hour course on how to learn about CrewAI, and all you gotta do is go to learn.crewai.com. And final thing, I promise. I know that we are right at time.
- 11:44
Bear with me. This conference has been one of the best conference I have been. Who here agrees with that?
- 11:50
Whoo. [claps]
- 11:52
Right? There's one thing, though. I heard a lot of, like, theory, a lot of, like, coming soon type of news, and I don't know about you, but I'm a little sick of that.
- 12:05
So why don't we actually start to bring some, like, agents into production?
- 12:11
So I wanna talk about CrewAI Plus. It's our enterprise offering, what some of those companies that I showed you were, are using it. And with CrewAI Plus, now you build your crews the way that they're running our terminal, but you basically can select them, push them to GitHub.
- 12:24
In three minutes, they become an API, and I'm talking about a real API. I'm talking about auto-scaling, protected by a bearer token with a private VPC, everything that you need to run these things in production.
- 12:36
And then you can also, like, one click away, export that into a React component, and now you basically have a UI that you can demo, and you can customize it any way that you want.
- 12:44
So you can basically connect your agents, like, in a few minutes to anything.
- 12:49
And there is more. For the first 50 companies that sign up using this link, we're gonna give you access to CrewAI Plus in less than 24 hours. And I also got one extra thing.
- 13:01
You don't even have to build your first crew. Turns out I pulled an all-nighter. I have a crew that can build your crew. [laughs]
- 13:09
You heard me right. Based on your email and your company name alone, this crew is gonna run. It's gonna create your crew, push into a GitHub repository, and that can be the first crew that you deploy on CrewAI Plus.
- 13:19
So hey, why don't we leave at that? It's starting to become a little, like, uh, ugly for the other guys. So hey, thank you so much. I catch you in conference. [claps] [upbeat music]