AI Engineer World's Fair 2024

Using agents to build an agent company

João Moura· CEO, crewAI13:46

Loading slides…

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.

Diagram linking LLMs to next-token prediction and reaction, ending with “user stop being a blocker” and “an agent is born.”Open full source frame
From LLM content generation to an autonomous agent.
0:000:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:00 · section reference included

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.

ScopeAdditional concerns
One agentCaching, memory, training, guardrails
A crew of agentsCommunication, shared cache, shared memory
Multiple crewsCoordination 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.

1:101:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:10 · section reference included

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.

4:364:49
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:36 · section reference included

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:

  1. Analyze a lead’s responses.
  2. Compare those answers with CRM data.
  3. Research the lead’s industry.
  4. 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.

6:356:44
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:35 · section reference included

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.

Slide showing Carlyle, PwC, K2, KPMG, Aurecon, Capgemini, and Havas logos.Open full source frame
Company logos presented during the adoption discussion.

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.

8:328:41
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:32 · section reference included

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.

Agent code example with role, goal, backstory, and allow_code_execution=True, with the execution flag underlined in red.Open full source frame
The agent configuration highlights allow_code_execution=True.
9:4810:04
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:48 · section reference included

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.

10:1310:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:13 · section reference included

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.

11:5212:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:52 · section reference included

Resources

From the talk

Updates since the talk

Transcript download
Full transcript

Loading transcript…