Key Moments
AI Dev 26 x SF | João Moura: Building Recurring, Governed, and Embedded Enterprise Workflows
DeepLearning.AIWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Enterprises are grappling with AI operationalization, moving beyond simple automations to complex, recurring workflows. CrewAI's CEO highlights that while building AI is commoditized, the real value lies in creating auditable, scalable, and integrated systems that drive business outcomes.
Key Insights
Iris, an internal coding agent at CrewAI, initially faced user attempts to "break it" but evolved to alter almost half of the company's pull requests in a week.
CrewAI's user telemetry shows a significant and increasing number of agent executions beyond coding, indicating a broader adoption of AI for business-related use cases.
The distinction between ad hoc (output-focused) and embedded (process-focused) agentic workflows is blurring, with users expecting conversational interfaces and interconnected agent tasks.
Reusable building blocks for tools and agents, including open-source skills available at skills.creai.com, are crucial for accelerating AI adoption and fostering organizational reuse.
Successful large-scale AI adoption in enterprises hinges on having a clear strategy for what to build first and next, and on enabling broader user adoption of these solutions.
The future of AI development likely involves self-improving, long-running, and conversational agents, emphasizing an 'entangled' agent ecosystem that learns and evolves with use.
The rapid evolution and adoption of AI agents
The industry is moving at an astonishing pace, making it challenging to keep up. CrewAI's CEO, João Moura, shared insights into how enterprises are tackling this by moving beyond simple AI experiments to building recurring, governed, and embedded workflows. He used the example of Iris, an internal coding agent at CrewAI, which initially faced resistance as users tried to "break" it by submitting unusual requests. However, Iris evolved significantly, to the point where it now alters nearly half of the company's pull requests weekly. This anecdote illustrates the potential of AI agents, moving from initial skepticism to widespread adoption and critical function within the organization. The broader adoption is evident in CrewAI's telemetry data, which shows a massive increase in agent executions each quarter, spanning well beyond coding to encompass various business-related use cases. This trend signifies a shift towards integrating AI more deeply into core business operations.
Beyond coding: Emerging AI use cases
While coding remains a prevalent use case for AI agents, the scope is rapidly expanding. Moura highlighted an internal use case where AI transforms sales meeting notes into personalized leave-behind materials for potential customers. This example is significant because it addresses tasks that are impractical to scale manually; creating custom materials for every client is resource-intensive. Through AI, such personalized outputs become feasible, demonstrating AI's capability to enhance customer engagement and provide tangible business value. This shift signifies that AI is not just about accelerating existing processes but also about enabling entirely new capabilities that were previously out of reach for many organizations.
The blurring lines between ad hoc and embedded workflows
Moura outlined two primary types of agentic systems: ad hoc workflows, where the output is paramount and the process disposable (e.g., generating a spreadsheet), and embedded workflows, where the process is as critical as the output (e.g., a healthcare system's doctor approval process). Historically, these were seen as distinct, with different tools catering to each. However, this delineation is rapidly blurring. Users now expect conversational interfaces for all agents, regardless of their purpose. Furthermore, the demand for agents to communicate and trigger actions in other agents is increasing, creating more complex, interconnected systems. This evolution means that the tools and frameworks for AI must become more flexible, capable of handling both simple, output-driven tasks and intricate, process-oriented workflows within a unified environment.
Commoditization of building and the rise of reusable components
The act of building AI agents is becoming increasingly commoditized, meaning the real value is shifting. Moura emphasized the importance of "flywheels" – components that can be reused by humans, agents, and various systems. This concept extends to tool repositories, which should accommodate diverse integrations (e.g., OAuth, MCP, HOA) and be accessible across the organization for both code and no-code users. CrewAI's approach includes enabling the interoperability of different agent frameworks (LangGraph, 8K, Salesforce, ServiceNow) and promoting shareable, open-source skills (accessible via skills.creai.com). These reusable building blocks are crucial for accelerating adoption and unlocking new possibilities, especially when they can be leveraged by different user groups. For instance, a skill called 'decide' is embedded directly into engineers' terminals, encoding company decision-making logic to guide them when they are unsure, thus empowering faster, more aligned day-to-day choices.
Integrating humans and improving visibility
Bringing humans into the AI workflow is essential, and email notifications have proven to be a highly effective method within CrewAI. This approach reduces friction by allowing users to respond directly via email, which agents then process. This mechanism not only facilitates human oversight but also makes non-technical personnel more comfortable engaging with AI systems. Crucially, effective AI operationalization requires robust metrics. Organizations need to "zoom out" to monitor overall system health, cost, and execution numbers, and "zoom in" to debug individual traces and understand agent decision-making processes. The ability to perform both granular and high-level monitoring is vital for assessing performance, identifying issues, and ensuring the AI systems are functioning as intended and delivering value.
The strategic imperative: Discovery and adoption
A key learning for CrewAI was that adopting AI is not solely an engineering problem but a transformational one, especially for large enterprises. The companies succeeding with AI at scale (running millions of agents) possess a clear strategy: they know precisely which use cases to prioritize and how to drive adoption among their workforce. This insight led CrewAI to focus on the "discovery" piece – helping clients identify the most impactful use cases. For non-tech companies, understanding where to start is paramount. By providing data-driven guidance on optimal use cases, CrewAI helps organizations build a foundation for successful AI implementation. Closing the loop between discovery, building, and adoption, leading to self-evolving systems that create organizational memory, is where true "magic" happens.
Future trends: Self-improving and entangled agents
Looking ahead, CrewAI is investing in several key areas for future AI development. The emphasis is on self-improving agents that can run for extended periods without constant human prompting. Conversational interaction is becoming a first-class priority, applicable to both ad hoc and embedded workflows. The concept of "entangled agents"—agents that become more effective with increased usage and can create automatic memories, track decisions, and understand their provenance—is a significant focus. This evolution suggests a future where agents are deeply integrated, continuously learn, and adapt, fundamentally changing how software is built and operated. Traditional engineering principles like DRY (Don't Repeat Yourself) and graceful degradation remain highly relevant in this new paradigm.
Navigating the AI revolution: Control and agency
The rapid advancement of AI is causing significant disruption, particularly in the software industry, leading to a paradigm shift where software is "organically forming" through prompts and data. João Moura acknowledges the inherent lack of control individuals have over these broad technological shifts. However, he stresses that individuals retain control over how they choose to engage with these changes. The "good news" is in personal agency: learning to use these new tools and understanding how to deploy them will invariably put individuals in a better position, regardless of the uncontrollable external factors. His core advice to engineers is to proactively experiment with and adopt these emerging AI tools, as this direct engagement is the most effective way to navigate and benefit from the ongoing AI revolution.
Mentioned in This Episode
●Software & Apps
●Companies
Key Principles for Adopting AI Agents
Practical takeaways from this episode
Do This
Avoid This
Common Questions
CrewAI is a framework for orchestrating autonomous AI agents. While initially popular for coding tasks, its usage has expanded significantly into various business-related workflows, enabling companies to automate complex processes.
Topics
Mentioned in this video
A framework for orchestrating autonomous AI agents. The speaker discusses its internal use and broader adoption for various workflows.
An internal coding agent developed at CrewAI that uses CrewAI behind the scenes. It has evolved to be self-made, with its own memory, skills, and flows.
A messaging platform used internally at CrewAI, where the speaker announced the Iris project and where team members initially tried to break it.
An AI-powered code editor, mentioned as a popular tool for agents in coding.
A platform or tool for coding, mentioned as a competitor or alternative to Cursor and Codex in the context of AI agents for development.
An AI model from OpenAI, mentioned as a tool for AI-assisted coding, alongside Cursor and Cloud Code.
A framework for building stateful, multi-agent applications, mentioned as an example of an agent type that can be integrated with CrewAI.
A website where users can download open-source skills for creating and integrating agents, demonstrating a community-driven approach.
A large enterprise customer mentioned as an example of a massive company using CrewAI's solutions, indicating the scale of their operations.
A large enterprise customer mentioned as an example of a massive company that utilizes CrewAI's services, highlighting the reach of their solutions.
Mentioned in the context of model providers and how AI is commoditizing building, although not directly discussed as a product in this segment.
A CRM platform mentioned as a type of agent that can be integrated and work with other agents through CrewAI.
A platform for digital workflows, mentioned as an example of an agent type that can be integrated with CrewAI.
More from DeepLearningAI
View all 98 summaries
27 minAI Dev 26 x SF | Diamond Bishop: The Next 100 Agents. Building the Agent Native Office
22 minAI Dev 26 x SF | Andrew K. Davies: Deterministic Memory: How to Build an AI That Cannot Lie
25 minAI Dev 26 x SF | Ara Khan: Evals Are Broken Use Them Anyway
26 minAI Dev 26 x SF | Manos Koukoumidis & Stefan Webb: VibeML: Build your AI model in hours, not months
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free