#402 Max: The AI Workforce (GStack, Hermes, Superpowers, and Paperclip)

Fiscally responsible, financial geniuses, monetary magicians. These are things people say about drivers who switch their car insurance to Progressive and save hundreds. Because Progressive offers discounts for paying in full, owning a home, and more. Plus, you can count on their great customer service to help when you need it, so your dollar goes a long way. Visit Progressive.com to see if you could save on car insurance. Progressive Casualty Insurance Company and Affiliates' potential savings will vary, not available in all states or situations. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge, and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trading value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. Imagine buying a state-of-the-art professional kitchen today. Oh yeah. You spend an absolute fortune on the equipment, and then you only use it to toast bread. Right, which sounds completely ridiculous out loud. It really does. But honestly, that is exactly how we use AI today. Welcome to our deep dive. Glad to be here. We are unpacking a really fascinating article by Max Ann today. It covers a major shift happening in the industry right now. A massive shift, yeah. We are moving away from simple AI chat interfaces completely. We are heading toward a fully autonomous AI workforce instead. It is a totally different way of working. And we will explore four specific open source blueprints to get there. We are looking at superpowers, G-Stack, Hermes Agent, and Paperclip. Okay, let's untack this. We really need to understand the baseline problem first. A regular AI assistant is purely reactive by its very nature. Right. You ask a direct question, and you get an immediate answer. Which works perfectly for simple, highly isolated daily tasks. You just want a quick recipe or maybe a code snippet. It is fast. It is undeniably very useful for those tiny micro tasks. Yeah. But real meaningful work is rarely that incredibly simple. Building a software feature requires intense planning and rigorous testing. Exactly. It requires maintaining memory across many different sequential work sessions. And this is exactly where single chat AI falls apart completely. The entire illusion of intelligence just breaks down under structural pressure. I always compare it to a talented but forgetful intern. Oh, that is a great way to frame it. They might be brilliant at writing a single block of code, but they lack the architectural scaffolding to manage their own memory. Right. They need exact, precise instructions for every single subsequent step. If we connect this to the bigger picture, the next massive productivity jump is a completely different paradigm entirely. It is not about building one marginally smarter foundational model. It is about coordinating multiple capabilities with a very rigid structure. I mean, I get that it forgets things over time, but why does the single chat model break down so quickly on big projects? Well, it comes down to how context windows actually operate mechanically. Okay. An isolated chat window just cannot hold overarching project goals. It predicts the next word based on the immediate previous text. So it just reacts blindly to the very last prompt you typed. Exactly. It has no true long-term plan. Right. Complex work needs memory and multi-step processes, not just instant answers. Beat. We need a fundamental shift in our daily approach. We really need to build guardrails around how the model operates. Which brings us to fixing the sloppy code generation first. But how do we actually force that discipline onto an LLM? This brings us directly to a project called Superpowers. Right. It was created by a very clever developer who goes by Obra. Superpowers adds real software engineering discipline directly to Claude Code. I still wrestle with prompt drift myself when asking AI for code to Sex Island. Yeah, it happens constantly. It just kind of wanders off the main architectural path entirely. We all do. That is the fundamental nature of the probabilistic piece. AI coding agents are incredibly fast right now, but that raw speed often brings hidden, deeply frustrating structural errors. Because they skip crucial structural tests just to save a little time. Exactly. They introduce weird, convoluted shortcuts that inevitably break things later on. They produce code that works perfectly fine for today, but it becomes totally impossible for a human to maintain by tomorrow. But Superpowers actually forces the AI to slow down completely. It really does. It demands a totally clean, isolated workspace before writing anything. It uses Git work trees to keep the main environment safe. Which are isolated folders for safely testing different code changes. Right. It forces the AI to brainstorm before writing actual logic. Then it writes a highly detailed, step-by-step implementation plan. The documentation explains this core goal in a very direct way. The plan must be incredibly detailed and logically robust. Mm-hmm. Even an inexperienced human engineer could follow it correctly. It essentially removes all the dangerous, hallucinatory guesswork. It enforces a strict, test-driven development cycle on the model. Right. Writing automated tests before writing the actual software code. Exactly. Humans honestly hate doing test-driven development because it feels slow. Oh, absolutely. But an AI doesn't have the capacity to feel bored. It will happily write a failing test and then fix it. And then it explicitly requests a thorough code review. It forces the branch to finish properly and cleanly. Right. It prevents the model from just abandoning a half-finished thought. I get that slowing it down sounds good in theory, but aren't we just stripping away its main speed advantage? Why is slowing the AI down actually an upgrade here? Because raw speed often introduces hidden vulnerabilities and sloppy logical shortcuts. Oh, I see. Slowing down forces the AI to formally verify its own logic first. It makes the final output actually trustworthy and structurally maintainable. Got it. Forcing a pause prevents sloppy, hard-to-maintain code later on. Ugh. Beat. So, we have highly disciplined coding mechanics established now. The model is writing clean, tested, and verifiable software code. But perfectly written code for a terrible idea is still terrible. How do we evaluate the actual ideas being coded? Well, we have to give the AI different hats to wear. And this is where G-Stack enters the conversation perfectly. It addresses the exact blind spot you just mentioned. Yeah. It was created by Y Combinator president Gary Tan. G-Stack gives the AI a roster of highly specific roles. You are building a one-person startup team, essentially. Your AI gets a bunch of different highly specialized job titles. We are talking CEO, engineering manager, and lead designer. It also acts as a QA lead and security officer. It follows a very strict and methodical sprint flow. Think, plan, build, review, test, ship, and finally reflect. What's fascinating here is the underlying mathematics of the prompt. When you assign a specific persona, you shift token probabilities. Right. A general agent gives loose, generalized, and very broad responses. A role-based setup forces highly focused, specific domain decisions. Because each role looks at the exact same problem from a different angle. The security officer agent strictly looks for potential injection flaws. The designer agent only cares about user experience and interaction flow. It stops the AI from rushing into an immediate, generic implementation. It forces it to pause and evaluate from multiple perspectives. Inside Cloud Code, you access this through simple slash commands. You just type slash cstack or slash office hours to trigger the roles. It guides the AI through deeply structured analytical steps. It prevents the model from trying to be everything it wants. I have to offer a massive warning right here, though. Do not ask it to build an entire app on day one. Well, definitely not. It is exactly like hiring a brand new project manager. You cannot demand a shipped product on their first afternoon. That usually leads to incredibly messy and logically broken results. Yeah. The context window gets overwhelmed by too many competing priorities. So start small. Evaluate a tiny market opportunity or reframe a basic product idea. Get a design review or run a simple debugging workflow. Does assigning a fake CEO title actually change the underlying code the AI writes? Yes. Because it radically changes the mathematical context of the prompt. Okay. The AI evaluates the coding problem through a completely different semantic lens. It prioritizes different logical metrics based on that specific assigned persona. Makes sense. Different roles force the AI to catch its own blind spots. Beat. We have disciplined coding and we have structured, specialized roles. But there is still a massive missing piece to this puzzle. Right. Mid-role sponsor read. We need to talk about what happens when you close your laptop. Usually the AI just forgets everything you just did together. The context window wipes clean and you start from absolute zero. Let's fix that. This brings us to a project called Hermes Agent. It was built by the brilliant team over at NUCE Research. Hermes is a completely self-improving AI agent framework. Most AI tools start totally fresh every single time you open them. Hermes is trying to be the exact opposite of that temporary paradigm. It creates specific skills from its past interactions with you. Yeah. It searches through all your past conversations seamlessly using vector databases. Which builds a much better long-term picture of how you actually work. It features a truly unified messaging gateway as well. You can connect it directly to Telegram or your company Slack. You can use Discord, WhatsApp, Signal, or the standard command line. So you do not have to constantly switch between five different browser windows. You interact with the exact same intelligent agent everywhere you go. Oh! Imagine it's scaling to remember your specific workflow across every single app you use. It changes the entire fundamental relationship you have with the machine. Yeah. You are no longer asking highly isolated randomly generated questions. Mm-hmm. You are developing a persistent system that organically evolves alongside you. But we need to talk about the actual onboarding reality here. We have to issue a quick, very serious security warning first. Definitely. It is always worth reviewing open source installation scripts very carefully. You must do this before piping them directly into your local shell. Taking a moment to inspect the script reduces unnecessary local risk. Because you are giving an autonomous agent access to your file system. And you really need to be incredibly patient with this specific tool. Do not expect magically strong personalized results on the very first day. It behaves way more like onboarding a real human teammate. You have to correct it when it makes a logical mistake. It is absolutely not an instant feature toggle you just casually flip. People always expect these memory tools to be instantly telepathic. How long does it realistically take for this self-improving loop to actually feel useful? It usually takes several weeks of consistent daily interaction to calibrate. Wow, weeks. Yeah. The system needs enough varied data to understand your specific workflow patterns. It cannot learn your deep, nuanced habits from just two short conversations. So it's true onboarding. It builds rich context through real use over time. Oh, beat. So we have disciplined role-playing agents equipped with persistent cross-platform memory. Let's put them all in one digital building and track the budget. We're talking about Paperclip now. Right. It is easily the most experimental and fragile project on this list. It orchestrates multiple specialized agents as if they were a real company. Yes. Paperclip throws a React front-end on top of all this underlying complexity. It gives you an actual visual dashboard for everything. So you aren't just staring at scrolling terminal outputs all day. Exactly. Inside this interface, agents take on major corporate roles. You have a CEO, a CMO, and a CTO actively working together. It includes a full organizational chart and a functional template. It has a ticketing system. It has strict governance controls and granular financial budget tracking. You can see exactly how much each specific generative task actually costs. Which is crucial. Here's where it gets really interesting. I look at all this corporate scaffolding and I have to ask, is this just corporate process theater applied to AI? That is a very fair and deeply necessary critical question. I have to give the honest warning straight from the source text. Okay. This is absolutely not a magic money machine at all. You will not wake up to sudden unexpected surprise profits. The project is still incredibly rough and highly experimental software. It is meant for those genuinely curious about multi-agent orchestration at scale. Its true purpose is high level coordination, not just blind reckless automation. Because as soon as you use multiple agents, massive structural problems appear. Responsibilities blur quickly and agents get stuck in infinite conversational loops. And when they get stuck in loops, EPI costs quietly skyrocket. Oh, they will happily burn through your open AI credits arguing with each other. Paperclip attempts to bring visible structure to that massive underlying complexity. It places everything in one single unified and highly trackable interface. Why do we even need a whole dashboard just to manage AI agents? Because multiple agents working simultaneously create massive unreadable chaos very quickly. Right. Goals drift rapidly without central oversight and very clear financial boundaries. A dashboard makes the invisible chaotic work of agents visible and trackable. Exactly. It brings visible coordination and budget tracking to multi-agent chaos. Beat. We should probably synthesize the ultimate pattern behind all these different tools. Let us step back and look at the actual big picture. We just covered four radically different open source software projects today. But they all share the exact same underlying philosophical approach. Every single one makes the exact same major technological bet. The future of AI is not a race for the smartest individual model. Right. It is about building a much better structure around how models work. G-Stack gives your general models incredibly clear structurally defined roles. Hermes gives them deep, persistent, and highly accessible cross-platform memory. Superpowers gives them reliable, highly disciplined software engineering process. And Paperclip gives them a massive, visible organizational management structure. The industry focus is shifting away from raw, generalized benchmark intelligence. It is moving toward consistent, highly predictable behavior in real world use. That shift might feel slightly less dramatic than a viral benchmark screenshot, but it is way more important for anyone actually building real things. We should probably help people figure out where to actually start. I want to explicitly lay out the recommended order of operations here. Definitely. I do not want you to lose your entire weekend to confusing documentation. Do not try to install all four on a Saturday afternoon. That is a truly great way to end up with absolutely nothing useful. First, you need to start with the superpowers framework. Do this if you already use Claude Code on a daily basis. It is the absolute fastest way to see immediate, practical coding value. The improvement to your daily development process is stark and visible right away. Second, you should try implementing the G-Stack role system. Add those structured, specialized roles to your newly disciplined coding process. It pairs naturally with superpowers once you understand the basic operational layers. Third, move on to the Hermes agent persistent framework. That gives you the powerful cross-platform memory you really need. It is absolutely perfect if you want an agent that runs scheduled background tasks. Finally, you can carefully explore the paperclip dashboard. Only do this if you are genuinely ready for experimental multi-agent orchestration. It is the most powerful conceptual framework but definitely the roughest tool. Yeah, save it for last. Following this specific order helps you see practical value very quickly. You gradually understand exactly how each structural layer contributes to the whole. The future probably won't look like one singular genius chat assistant. No, they will look exactly like a small orchestra of highly specialized systems. Each has a very clear role and a rigidly defined process. They will smoothly hand off complex work to the next digital player. The massive operational advantage goes to those who design these workflows early. If AI is rapidly becoming the orchestra of the modern workforce, what skills do you need to start developing today to become the conductor? That is a perfect profound question to leave them with today. Thanks for joining us on this deep dive. Stay curious. Out to your own music. Capital One's tech team isn't just talking about multi-agentic AI. They already deployed one. It's called Chat Concierge and it's simplifying car shopping. Using self-reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get pre-approved for financing, and estimate trading value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. Let's be honest. Buying cannabis shouldn't be complicated, sketchy, or low quality. That's why I want to tell you about mood.com. That's M-O-O-D dot com. Mood ships federally legal cannabis straight to your door. No medical card, no hassle. And here's the kicker. The quality is better than anything you'll find at your local dispensary. Yeah, I said it. Whether you're into edibles, concentrates, flower, or just looking to explore, you'll find it all at Mood. And it's not just the variety that makes them stand out. Every product is sourced from small American-owned family farms that care deeply about what they grow. It's cannabis you can trust, delivered discreetly, and ready to elevate your mood. And because you're a listener, you get 20% off your first order. Just head to mood.com. That's M-O-O-D dot com to get started. That's M-O-O-D dot com to get started. I've got a lot of food inside. But as much Koreans live here in peeing some books. Then it's rápido. Oh,轟 Jadi животogen and sit down with a quick deadline of waiting on what we have to drink.