I’ve been spending a lot of time testing different AI agent harnesses lately, and I’ve had a bit of an "aha" moment. We tend to look at these tools as totally different things. But underneath the hood? They are almost identical. Think about it like programming languages. Every language, whether it’s Python, TypeScript, or anything else, is basically just variables, classes, and functions arranged differently. Once you understand those building blocks, you can pick up any language fast. AI agents are the same. Every tool I’ve used basically comes down to three simple parts: 1. The Brain (System Prompts): This is the personality. It’s how you set the default behavior and tell the AI who it is and how it should act before it even starts. 2. The Workflow (Repeatable Logic): This is how you control the step-by-step process. It’s the "recipe" you give the AI so it does the same task the same way every time. 3. The Hands (Extensibility): This is how you let the AI touch the world, using MCP servers, hooks, or custom plugins to connect to your real files, databases, or APIs. Why does this matter? Because once you realize this, you stop caring about the hype cycle. You realize you don't need the "hottest new tool." You just need to master how to build those three things. If you can control the personality, the workflow, and the extensions, you can build a production-quality product with any of these tools. Am I oversimplifying this, or is this the mental map you use too? How are you setting up your "harness" right now? Let's discuss in the comments. 👇

Lately I've been using Codex a lot and decided to make a video on it. I walk through the four levels of integrating Codex into your workflow: - Level 1: Dialing in your memory context, custom instructions, and building sub-agents so you aren't stuck with vanilla, out-of-the-box outputs. - - Level 2: Hooking up skills, plugins, and MCP workflows (like the Superpowers plugin) to expand what the agent can actually execute. - - Level 3: Using the Playwright MCP to give your agent browser automation capabilities to perform tasks on your behalf. - - Level 4: Setting up scheduled, hands-free recurring tasks (I show exactly how I set it up to scrape trending GitHub repos and Product Hunt launches every day). What are your thoughts on Codex??

Most business owners are using AI wrong. They treat it like a faster Google... Ask a question. Copy the answer. Move on. That's not AI. That's search with extra steps! The real move is using it as a THINKING partner on decisions you're already making. I had a pricing call a few week ago. Three options on the table. Instead of asking Claude "what should I price this at," which is a garbage prompt, I dumped in everything... - Who the customer is. - What they paid before. - What the deal looked like. - What my gut was telling me. THEN... I told it to argue against my assumptions. Here's why that matters. Your brain already has a preferred answer before you start thinking. Confirmation bias is real. Most people use AI to confirm what they already believe. All that does is make you more confident in a decision you were going to make anyway. But when you tell AI to be the opposition, you get something you can't get from your own head; a second authority that's emotionally detached from the outcome. Guess what? I changed my pricing after that conversation and made a better deal. Stop asking AI for answers. Start asking it to challenge yours.

Lately I’ve been trying to improve my entire YouTube video creation workflow. Not just “come up with better ideas,” but the whole process: Idea → research → script notes → video structure → post-production → retrospective after publishing. The main problem I kept running into was context switching. So I built a small workflow web app to bring the whole video creation process into one place. Inside the app I can track: - Quarterly YouTube goals - My target audience - My value proposition - New video ideas - Active video projects - Core research findings - Script notes - Segment structure - Published videos - Post-upload retrospectives - Audience feedback I am excited to see how the retrospective flow goes. After a video is published, I can come back and review what worked, what didn’t, what the audience responded to, and what I should improve for the next upload. I don’t just want to publish more. I want the videos to actually get better. For the build, I used Next.js and built most of this as a background project while working on other things. I also used Superpowers, and honestly, it worked really well for this kind of workflow app because I already had a clear idea of what I wanted. Sometimes the best AI projects are small internal tools that remove friction from a workflow you already do every week. I may put this on GitHub after I clean up a few more things. What do you guys think: What workflow do you currently have that would be better if it was turned into a small web app? And where are you still bouncing between too many tools?

I just realized that I forgot to share this with you guys... Feel free to watch the video but if you just want the rundown you can skim this summary i had ai write for us based on the video transcript😉 If you are building with AI agents, you know how frustrating it is to be tethered to your desktop while waiting for an agent to finish a complex coding task. I just dropped a new video breaking down Paseo, an open-source, local-first multi-agent orchestration tool that completely solves this problem. It lets you monitor, control, and execute top-tier coding agents from an iOS or Android device. The Summary: What is Paseo? Paseo is a multi-agent unification platform designed for local-first software development. Utilizing a client-server architecture similar to Docker, Paseo runs a local daemon on your machine (supporting Windows, Linux/WSL, and macOS) that securely communicates with mobile clients. This setup allows developers to spin up a FastAPI server, review files, split terminal panes, and switch between models like Claude 3.5 Sonnet (Cursor), OpenCode, Pi Agent, and Codeex seamlessly from a smartphone. Key Technical Features & AI Architecture - Local-First Code Security: Your source code never leaves your local machine. Paseo functions entirely within your local development environment. - End-to-End Encrypted Relay: Remote access (e.g., controlling agents while away from your desk) relies on an untrusted relay server. Traffic between your phone and the local daemon is completely end-to-end encrypted, ensuring third parties cannot read your code or modify data packets. - Unified Agent Orchestration: Features a powerful split-pane UI where you can run a Cursor model on one side, an OpenCode agent on another, and an active terminal tab simultaneously. - Global System Prompts: Streamlines agent engineering by allowing you to pin a single, universal system prompt across all connected AI providers.