AI agents can replace brittle, predefined automation flows by choosing tools and adapting to changing circumstances.
2
Joao Moura used crews for marketing, lead qualification, documentation, and email work as crewAI grew.
3
Teams should start with simple, low-risk, high-impact agent use cases, then move agents into production with the right operational controls.
Summary
Joao Moura explains how AI agents differ from traditional software automation. Conventional systems depend on known inputs, fixed logic, and predictable outputs. Agents can choose actions and tools at runtime, which makes new kinds of automation possible but also introduces memory, caching, training, guard rails, and coordination problems. Moura tells the story of using crewAI to build an agent-based company. He started with a marketing crew, then added lead qualification, documentation, and email workflows. The marketing system produced more views, while the lead qualification workflow helped him handle a large increase in customer calls. He argues that companies should begin with simple, low-risk, high-impact tasks. The talk also introduces crewAI features for code execution, crew training, third-party agents, and enterprise deployment through APIs, private VPCs, and React components. His central argument is practical: agents are already being used at scale, so builders should start applying them to real work.
Agents can adapt where fixed automation becomes tangled
Moura contrasts agents with the automation engineers have built for decades. A normal workflow starts with a known path from A to B, then becomes harder to maintain as more conditions and steps are added. With agents, the developer can provide a set of options and let the system adapt to the circumstances it encounters in real time. This makes some automations possible even when every input and branch cannot be specified in advance. The tradeoff is that the system has less predictable behavior than traditional software.
Production agents need much more than a model and a tool
A simple agent may appear to contain an LLM, a task, and some tools. Production systems quickly require more infrastructure. Moura names caching, memory, training, and guard rails as parts of the design. When agents work together, memory and caching may need to be shared across the group. Multiple crews communicating with one another add another layer of coordination. The architecture becomes more involved as an agent moves from a demo into a system that has to run repeatedly.
AI applications require a different testing model from deterministic software
Traditional software usually has typed inputs, explicit operations, and a predictable output, so engineers can write tests around stable behavior. AI applications are fuzzy. An input described as a string could contain a CSV, a rasp, or a random joke, while the model behaves like a black box and the output is harder to predict. Moura says he likes this change, but it means engineers need to think differently about how they build and evaluate software.
crewAI grew through real usage and community demand
Moura describes crewAI as a production-ready library for building and orchestrating multi-agent automations. He says more than 100,000 crews were executed each day and more than 10 million agents had run in the previous 30 days. The project also gained more than 16,000 GitHub stars, a Discord community of more than 8,000 people, and a Reddit community. As engineers and companies began reaching out, the scale of interest forced the team to think about how to grow the company.
Moura used crews inside the company before expanding their scope
The first company workflow he describes is a marketing crew made up of a content creator specialist, social media analyst, senior content writer, and chief content officer. The crew checked social networks, searched the internet, and considered Moura's previous experience before turning rough ideas into drafts. He says this produced ten times more views over a 60-day period. He then built a lead qualification crew with a lead analyst, industry researcher, and strategic planner. It analyzed lead responses, compared them with CRM data, researched the industry, and produced scores and talking points for meetings.
Starting with low-risk, high-impact work gave the company room to learn
After marketing and lead qualification, Moura expanded crews into documentation and email. He presents this sequence as a way to adopt agents inside a growing company without beginning with the most dangerous or complicated work. His advice is to start simple, expand to use cases with low risk and high impact, and learn from the results. The examples are internal workflows that reduce repetitive work while keeping Moura involved in customer calls and decisions.
crewAI is adding code execution, training, and outside agents
Moura announces several crewAI features. A code execution option allows agents to write and run code through an automated coder and command-line code executor. A new crew training CLI lets users give instructions that become part of the agents' memory, with the goal of producing more consistent results over time. He also says crewAI will accept agents from other frameworks, including Yama index, Link sh, and AutoGen, while giving them crewAI features such as shared memory and common tools.
crewAI Plus packages crews for production deployment
The enterprise offering, crewAI Plus, lets users select crews from a terminal workflow and push them to GitHub. Moura says a crew can become an API in three minutes, with autoscaling, bearer-token protection, and a private VPC. Users can also export a crew into a React component to create and customize a user interface. He offers a crew that generates another crew from an email address and company name, then pushes the result to a GitHub repository for deployment on crewAI Plus.
"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."01:42
Who should watch
You are deciding where agents could replace repetitive internal workflows and need examples from a company that started with marketing and lead qualification.
You are building multi-agent systems and want to understand why memory, caching, training, guard rails, and coordination appear once agents move into production.
You are evaluating crewAI and want to hear about its code execution, crew training, third-party agent support, and enterprise deployment features.
