Dec 28, 2024 · 1h 42m · startup-ideas
How To Create Ai Agents From Scratch (CrewAI, Zapier, Cursor)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this hands-on masterclass, CrewAI co-founder and CEO João Moura joins Greg Isenberg to demonstrate how to build, debug, and deploy production-ready multi-agent AI systems from scratch. The tutorial covers end-to-end inbound lead enrichment, dynamic PDF report generation, and complex multi-crew orchestration using CrewAI Flows, Zapier, and Cursor.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 8.5% of the talking time here. How this is scored →
Greg as informed peer 2.4 Guest teaching 4.4 Guest disagreement 0.1 Greg pushing back 0.3
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:35:55 Critiquing the niche focus of generated content
João criticizes the agent's unexpected focus on video game SEO rather than general SEO with game examples, though Greg quickly spins it positively.
Hardest push from Greg ▶ 1:36:14 Reframing output anomaly into a business nicheGreg pushes back against João's view that the agent output was flawed, arguing that a dedicated gaming SEO curriculum is actually a high-value untapped niche.
Biggest teaching moment ▶ 13:28 Explaining role prompting and model benchmarkingJoão clearly educates Greg on how role definitions alter LLM outputs and explains how CrewAI Test automates model benchmarking against hallucinations and latency.
Greg holds their own ▶ 37:48 Demonstrating practical developer tooling perspectiveGreg demonstrates technical fluency by explaining his CS background and comparing rapid web prototyping in Bolt with AI editors like Windsurf, Cursor, and Lovable.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Designing Lead Enrichment Workflows in CrewAI Studio | 1 | 5 | 0 | 0 | João takes the lead screen-sharing and walking through the architecture of a lead enrichment workflow in CrewAI Studio, while Greg listens passively. | |
| Sponsor Break: Startup Empire Community | 0 | 0 | 0 | 0 | Solo host ad read for Startup Empire Community followed by João resuming the setup of CrewAI studio agents. | |
| Benchmarking LLMs and Evaluating Role-Based System Prompts | 4 | 7 | 1 | 2 | Greg asks pointed questions about model selection and whether persona descriptors like 'Senior' meaningfully change LLM behavior. João educates him on empirical benchmarking using the CrewAI Test CLI tool and role steering. | |
| Local CLI Debugging and GitHub API Deployment | 2 | 4 | 0 | 0 | João demonstrates local CLI installation and debugs an unassigned agent before pushing the code to GitHub for automated deployment. | |
| Connecting Webflow, Zapier Webhooks, and Resend | 4 | 2 | 0 | 1 | Greg probes on why João selected Webflow over Framer and provides domain context explaining Resend's transactional email architecture. | |
| Sponsor Break: BoringMarketing.com SEO Agency | 3 | 4 | 0 | 0 | After an ad read, Greg highlights the business value proposition of replacing $50k-$80k junior marketing report generation with automated agents, while João tests the Zapier webhook. | |
| Designing PDF Templates and Pydantic Data Structures | 4 | 3 | 0 | 0 | João queries Greg about his developer setup (Cursor vs. Windsurf), prompting Greg to share his CS background and practical preference for Bolt and Lovable. | |
| Refactoring Agent Configurations for Structured Pydantic Output | 3 | 4 | 0 | 0 | João refactors agent YAMLs using Cursor to output structured Pydantic models. Greg observes that getting a minimum viable output running before optimizing prompts is the right tactical approach. | |
| Writing Deterministic After-Kickoff Hooks for PDF Compilation | 1 | 6 | 0 | 0 | João walks through adding deterministic `@after_kickoff` hooks to inject structured Pydantic data directly into an HTML template for PDF compilation. | |
| Introduction to Stateful Orchestration with CrewAI Flows | 2 | 6 | 0 | 0 | João introduces CrewAI Flows, contrasting basic linear agent execution with event-driven, stateful multi-crew orchestration for complex long-form content generation. | |
| Building a Planning Crew and Plotting Flow Architecture | 2 | 5 | 0 | 0 | João codes a content planning crew and runs the CLI plot command to render a visual state diagram, which Greg enthusiastically praises as essential for visual thinkers. | |
| Chaining Multi-Agent Chapter Generation in a Loop | 1 | 5 | 0 | 0 | João implements a loop iterating through generated chapters to invoke dedicated chapter-writing crews equipped with Serper and web scraping tools. | |
| Live Execution and Debugging of the Course Generation Flow | 2 | 4 | 0 | 0 | During live execution, a string attribute error triggers live debugging where João adjusts the flow to return typed Pydantic objects instead of raw JSON strings. | |
| Resolving Pydantic Schema Errors and Configuring PDF.co in Zapier | 2 | 5 | 0 | 0 | João connects PDF.co in Zapier for dynamic rendering and resolves another live coding error where Cursor mismatched agent names and Pydantic attribute accessors. | |
| Generating Comprehensive SEO Course Content via CrewAI Flow | 5 | 4 | 1 | 1 | João runs the Flow with Greg's custom SEO and gaming parameters. When João notes the model's gaming SEO slant was unexpected, Greg reframes it as a viable niche market opportunity. | |
| Prompt Optimization, Guardrails, and AWS Infrastructure Crew Demo | 3 | 6 | 0 | 0 | João showcases how to separate fuzzy agent prompts in YAML from deterministic guardrails and code checks in Python, concluding with an AWS cloud monitoring demo and course recommendations. |
Statements from this episode (18)
Moura: CrewAI scaled its operations using its own AI agents
“From the get go, being able to use AI agents to scale an AI agent company was kind of like an interesting kind of like inception like experience.”
Moura: CrewAI agents do not require a 1:1 agent-to-task mapping
“With AI agents and crew AI specifically, you don't need to have one agent to one task. It's not necessarily one to one.”
Moura: OpenAI's o1-mini struggles with agents by avoiding external tools
“Honestly, O one mini in my experience for agentic behaviors, sometimes falls short. Just because it already has kind of like chain of thought embedded on it. It tries to basically like Figure out everything itself without using any tools and kind of like outsm…”
Moura: 7B and 14B parameter models loop excessively as AI agents
“Smaller models, like smaller, or seven B, 14 B, they don't work as well. You can get them to behave as agents, but they usually take even longer because they go in so many blind alleys and keep like retrying things. So you definitely wants to go, like my go-to…”
Moura: Role framing in system prompts significantly changes LLM agent output
“Because these agents, they're kind of like impersonating roles. You do get different behaviors depending on, you do get different behaviors depending on how you give them like a different role. So you can actually replicate this on ChatGPT. Like if you ask Cha…”
Moura: CrewAI Enterprise Automatically Converts Workflows Into APIs
“The thing is once that you're using the enterprise that automatically becomes an API.”
Moura: CrewAI Enables Direct Export of React UI Components
“Or if you want to, you can also export a UI. You can export a react component right from here.”
Moura: Email feels too consolidated to be an obvious startup idea
“It's funny when you think about email, like it's not kind of like your go to, you can like startup if you're trying to think about starting a business feels like such a consolidated market.”
Isenberg: Hiring a junior marketer for custom reports costs $50k–$80k
“This is something that you would hire typically a junior marketing person to actually put together, take the data, make it look nice. Maybe you want a designer there or whatever, but it's someone you're paying, you know, in the U S at least kind of, you know, …”
Moura: B2B marketing collateral should empower internal champions in meetings
“One thing that I would think is I would love to send a lead materials that they would like to show in a meeting. Right? Like imagine saying something so beautiful that they are okay taking that and going back to their team and show like, oh, let me show you wh…”
Isenberg: Bolt caters to non-technical users more than Cursor
“And I found that Bolt has been more geared towards non-technical people.”
Moura: Complex agent use cases require deterministic hooks for consistency
“As the use case has gotten more complex, you start to need more precision in more like a repeatability and consistency on their use cases. So usually people start with before and after hooks.”
Moura: CrewAI flow plot generates visual architecture diagrams for agent workflows
“You can do now crew AI flow plot. And that will actually plot you a visual showing exactly what is happening with your flow. So you can see what are the crews that you have in there. You can see what are the inputs, like what are the different kind of like fun…”
João Moura: Cursor Will Not Replace Software Engineers Anytime Soon
“Again, silly me thinking that cursor is gonna replace engineer so soon. It's gonna take quite a while.”
Moura: AI coding agents accelerate 0-to-1, but 1-to-10 remains manual
“Definitely AI and agents and things like Devon now and like other solutions, they help a lot, but it's still like a lot of like, just making sure that from going, going from zero to one is very fast and getting faster and faster. Going one to 10 is still very …”
Moura: Agent behavior is tuned in YAML, while precision is enforced in Python
“A lot of like the optimization, like what happens, especially on the fuzziness parts, like on the actual agents is on these agents and tasks. I got a YAML files. It's where I can, like, you're going to be like spending a lot of time, like to make sure that you…”