Building & Shipping 11 guests | 25 insights

Building With AI Agents

Transition from writing lines of code to directing a parallel team of autonomous agents.

Includes our free 8-email course on making your product team AI-native. Unsubscribe anytime.

The Guide

5 key steps synthesized from 11 experts.

1

Audit and Delegate

Identify repetitive tasks using the junior intern model. Scope these tasks to specific sub-tasks that require basic judgment but are energy-draining for humans, such as syncing sources of truth or basic refactors.

Featured guest perspectives
"Working with Devin is familiar because it feels like adding a few junior engineers to your team. You toss them tasks, and they’ll get started with enthusiasm . . . It’s programming leverage—it’s productivity power."
— Lenny Rachitsky
"Ask yourself: What ongoing work requires some judgment and writing abilities—but not your full expertise and intuition? Put another way, if my company assigned me a junior intern, what would I have them do?"
— Lenny Rachitsky
"Just as when delegating to people, the fanciest AI will perform only as well as the instructions it’s given. Are you clear on how you would accomplish this task manually, with mouse, keyboard, and coffee?"
— Lenny Rachitsky
2

Contextualize the Workspace

Create plain-text markdown files throughout your codebase to explain directory structures and navigation to any agent. Provide specific templates and examples of past successful work to guide the AI's output format.

Featured guest perspectives
"Just as when delegating to people, the fanciest AI will perform only as well as the instructions it’s given. The best way to gain this clarity is to do the task once or twice."
— Lenny Rachitsky
"I'm not sure who he's quoting, but code is just words at the end of the day. So it's just files on your computer. So basically you can be working on the same project and carry it from app to app. And especially now, I can work with multiple models and apps on my project."
— Zevi Arnovitz
3

Orchestrate Parallel Agents

Integrate agents directly into your communication tools like Slack or Linear. Assign distinct end-to-end tasks to multiple agent threads simultaneously, moving from linear development to parallel execution.

Featured guest perspectives
"What Devin does is it is a full asynchronous workflow, and so you can tag Devin on an issue in Slack, you're talking about an issue and you tag Devin, you can tag Devin in Linear, you can have Devin and Devin will make pull requests in your GitHub, and so it's very much built to work with engineering teams as your junior engineer."
— Scott Wu
"Our whole team is only like 15 engineers a year. We use a ton of Devin when we're building Devin. Most folks on the team are definitely working with up to five Devins at once, and so Devin merges like several hundred pull requests into production in the Devin code bases every month."
— Scott Wu
"Engineers are becoming tech leads. They're managing fleets and fleets of agents. It literally feels like we're wizards casting all these spells and these spells are kind of like going out and doing things for you."
— Sherwin Wu V2
4

Debug Logic over Syntax

When troubleshooting, prioritize reading the agent's natural language explanations over the code itself. If an error occurs, ask the AI to reflect on why its instructions failed and update your system prompts immediately.

Featured guest perspectives
"In order for you to steer the ship, you have to know the instructions, right? And the best way to learn is by building, but treating these tools almost as technical co-founders and educators, and learning while doing, and religiously reading the agent output."
— Lazar Jovanovic
"Explaining exactly what you expect and what you're not getting is even more important with AI than with the humans."
— Anton Osika
"Zevi shows you how to work with cursor to quickly add your ideas to Linear to explore your idea with AI, how to develop your plan, how to then build the thing, and then have different LLMs review your code and update your documentation, and then use all of this as a learning opportunity to develop your own sense of how things work."
— Zevi Arnovitz
5

Implement Automated Multi-Step Reviews

Structure your workflow so agents first analyze requirements before executing. Use a second AI model to peer-review the first model's logic before any pull request is merged into the main codebase.

Featured guest perspectives
"It's very difficult for me to catch mistakes. What I'll do is basically /review. This tells Claude to start reviewing its own code, but what's even cooler is I have Codex as well as Cursor open. I will have each of them review the code."
— Zevi Arnovitz
"So we have a Codex code review that's catching a lot of mistakes. It's actually caught some pretty interesting configuration mistakes."
— Alexander Embiricos

Get this guide as an AI skill for Claude Code

Includes our free 8-email course on making your product team AI-native. Unsubscribe anytime.

Install This Skill

Add this skill to Claude Code, Cursor, or any AI coding assistant that supports Agent Skills.

Quick Install (Recommended)

Install this skill directly using npx:

npx skills add RefoundAI/lenny-skills --skill building-with-ai-agents

Or install all 76 skills:

npx skills add RefoundAI/lenny-skills
View on GitHub →
Manual Installation
1

Download the skill

Download Skill (.zip)
2

Add to your project

Create a folder in your project root and add the skill file:

.claude/skills/building-with-ai-agents/SKILL.md
3

Start using it

Claude will automatically detect and use the skill when relevant. You can also invoke it directly:

Help me with building with ai agents

Guest Perspectives

Deep dive into what 10 podcast guests shared about building with ai agents.

Alexander Embiricos 1 quote
Listen to episode →
"So we have a Codex code review that's catching a lot of mistakes. It's actually caught some pretty interesting configuration mistakes."
Tactical:
  • Deploy AI agents to review their own infrastructure and training code to catch configuration errors.
  • Build tools that prioritize making it easier for humans to review code rather than just writing it.
  • Use AI to provide automated on-call support for critical infrastructure and training runs.
View all skills from Alexander Embiricos →
Anton Osika 1 quote
Listen to episode →
"Explaining exactly what you expect and what you're not getting is even more important with AI than with the humans."
Tactical:
  • Specify the exact text, behavior, or UI component that needs correction rather than giving vague instructions.
  • Use visual selectors to isolate problematic elements before requesting a change.
  • Be more descriptive with your AI agent than you would be with a human to ensure it understands the nuance of your request.
View all skills from Anton Osika →
Boris Cherny 1 quote
"100% of my code is written by Claude Code. I have not edited a single line by hand since November. Every day, I ship 10, 20, 30 pull requests. So, at the moment I have, like, five agents running."
Tactical:
  • Direct multiple AI agents simultaneously to handle different pull requests.
  • Stop manual code editing to focus on high-level direction and review.
  • Leverage AI tools to handle implementation, testing, and shipping minutia.
View all skills from Boris Cherny →
Dhanji R. Prasanna 1 quote
Listen to episode →
"What's been surprising and really amazing, the non-technical people using AI agents and programming tools to build things, the people that are able to embrace it to optimize for their particular workday and their particular set of tasks are really showing the most impact from these tools."
Tactical:
  • Provide non-technical teams with access to AI agents and programming tools to automate their unique workflows.
  • Encourage employees to use AI to optimize their specific workday and repetitive manual tasks.
  • Promote the use of AI agents that work autonomously during downtime to build in anticipation of upcoming team needs.
View all skills from Dhanji R. Prasanna →
Guillermo Rauch 1 quote
Listen to episode →
"But there's a little bit of a writer's block sometimes. So one of my favorite things that I've seen, and I'm even looking at the home page right now, and you can see a random assortment of community submissions. And they have 1,200 forks, and 1,500 forks, and 6,000 forks, and this is every time people saying like, 'Oh, instead of starting from scratch, I'll start from this application that someone else has built and I'm going to prompt it to modify it and make it my own.'"
Tactical:
  • Browse community galleries for starting points like menus, logos, or app layouts.
  • Fork existing prototypes and prompt the AI to modify them to fit your specific use case.
  • Invert the workflow by starting with high-level intent prompts to generate the first version of code automatically.
View all skills from Guillermo Rauch →
Lazar Jovanovic 2 quotes
Listen to episode →
"AI just don't understand what do you mean when you say, 'You know what I mean?' So you need to be specific. I'm optimizing 100% of my time today on good judgment, clarity, quality, taste."
Tactical:
  • Avoid vague phrases like 'you know what I mean' when prompting an AI agent.
  • Treat AI tools as technical co-founders that require clear, granular instructions to succeed.
  • Optimize for clarity in the 'ask' rather than focusing on the speed of the code generation.
"In order for you to steer the ship, you have to know the instructions, right? And the best way to learn is by building, but treating these tools almost as technical co-founders and educators, and learning while doing, and religiously reading the agent output."
Tactical:
  • Religiously read the agent's natural language explanations to understand the logic of the generated code.
  • Trust the AI's ability to handle syntax while you focus on steering the high-level intent.
  • Use chat mode to treat the AI as an educator when you encounter technical blocks or errors.
View all skills from Lazar Jovanovic →
Marc Andreessen 1 quote
Listen to episode →
"Over the holiday break, it feels like the AI coding thing really hit critical mass and the world's best programmers, including Linus Torvalds, for the first time over the holiday break basically said, 'Yeah, AI is now coding better than we can.'"
Tactical:
  • Use AI tools to transition from being a standard programmer to a super-empowered individual with 10x output.
  • Harness AI reasoning for complex tasks in coding, science, and math where answers are verifiable.
  • Adopt AI coding tools immediately to stay competitive with the world's best programmers who have already hit critical mass.
View all skills from Marc Andreessen →
Scott Wu 3 quotes
Listen to episode →
"Devin is a fully autonomous software engineer that is going to work on tasks end to end, and so there are a lot of great tools for all parts of the stack of the AI code workflow. What Devin does is it is a full asynchronous workflow, and so you can tag Devin on an issue in Slack, you're talking about an issue and you tag Devin, you can tag Devin in Linear, you can have Devin and Devin will make pull requests in your GitHub, and so it's very much built to work with engineering teams as your junior engineer."
Tactical:
  • Assign end-to-end tasks to agents via Slack or issue trackers like Linear.
  • Treat agents as asynchronous junior team members rather than just synchronous chatbots.
"What Devin does is it is a full asynchronous workflow, and so you can tag Devin on an issue in Slack, you're talking about an issue and you tag Devin, you can tag Devin in Linear, you can have Devin and Devin will make pull requests in your GitHub, and so it's very much built to work with engineering teams as your junior engineer."
Tactical:
  • Tag agents directly in Slack or Linear to initiate workflows without leaving team communication tools.
  • Ensure agents are integrated into GitHub to allow them to make pull requests automatically.
"Our whole team is only like 15 engineers a year. We use a ton of Devin when we're building Devin. Most folks on the team are definitely working with up to five Devins at once, and so Devin merges like several hundred pull requests into production in the Devin code bases every month."
Tactical:
  • Shift your mindset from synchronous single-tasking to asynchronous coordination of multiple agents.
  • Assign distinct tasks to different agent instances simultaneously to multiply your individual engineering throughput.
View all skills from Scott Wu →
Sherwin Wu V2 2 quotes
"Engineers are becoming tech leads. They're managing fleets and fleets of agents. It literally feels like we're wizards casting all these spells and these spells are kind of like going out and doing things for you."
Tactical:
  • Transition from writing code to managing multiple parallel agent threads simultaneously.
  • Treat programming languages as incantations used to steer models toward desired outcomes.
  • Focus your energy on steering agents and providing feedback rather than manual implementation.
"There's a team that's actually doing an experiment right now within OpenAI where they are maintaining a 100% Codex-written code base. They run into the exact problems that you're describing. And so usually you're like, 'All right, I'll roll up my sleeves and figure it out.' This team doesn't have that escape hatch."
Tactical:
  • Resist the urge to manually fix code when AI agents struggle or go off-rails.
  • Commit to maintaining portions of your codebase using only AI-generated outputs to build steering proficiency.
  • Develop seniority by learning to monitor and correct agents before their tasks diverge from the objective.
View all skills from Sherwin Wu V2 →
Zevi Arnovitz 3 quotes
Listen to episode →
"It's very difficult for me to catch mistakes. What I'll do is basically /review. This tells Claude to start reviewing its own code, but what's even cooler is I have Codex as well as Cursor open. I will have each of them review the code."
Tactical:
  • Create a /review command that forces the primary AI to check its own work for logic errors before deployment.
  • Keep a second AI model open (like Codex) to provide a 'second opinion' on complex code blocks.
  • Treat one model as the dev lead who must defend its code against the critiques raised by the other model.
"Zevi shows you how to work with cursor to quickly add your ideas to Linear to explore your idea with AI, how to develop your plan, how to then build the thing, and then have different LLMs review your code and update your documentation, and then use all of this as a learning opportunity to develop your own sense of how things work."
Tactical:
  • When the AI makes a mistake, ask it to reflect on what in its instructions caused the error.
  • Update your system prompts and custom /commands immediately after resolving a recurring technical bug.
  • Maintain a folder of shared knowledge in your project that the AI must consult before every new task.
"I'm not sure who he's quoting, but code is just words at the end of the day. So it's just files on your computer. So basically you can be working on the same project and carry it from app to app. And especially now, I can work with multiple models and apps on my project."
Tactical:
  • Add plain-text .md files throughout your codebase to guide AI agents on how to work in specific directories.
  • Keep high-level documentation updated so that any model (Claude, Codex, Gemini) can quickly gain context on the project.
  • Use file-based storage to easily move your project between different AI tools and leverage their unique strengths.
View all skills from Zevi Arnovitz →