# AI Mastermind & Mixer — Clean Transcript
**Date:** 2026-08-07 · **Type:** AI Mastermind (free monthly, first Friday)
**Location:** Industrious, Austin, TX
**Duration captured:** 2:21:01 · Main program begins ~[00:11:00]

> **Timestamp note:** times below are from the RAW recording (which started at
> noon, during the lunch mixer). Re-map them to the published/edited video
> before using any "watch at" links.
> **Attribution note:** triangulated from three captures — the ceiling-mic Zoom
> connection (Fathom), a Granola transcript, and the Plaud tabletop recorder
> with Austin's post-event speaker tagging (attribution ground truth) — plus
> the live idea board's submission records. `[H]` high confidence (Plaud voice
> tag, two sources agreeing, or host-named/self-introduced) · `[M]` verify
> speaker · `[V]` verify wording. Where we
> could not determine the speaker, the segment is marked **Unattributed**
> rather than guessed. This transcript is edited for signal: filler, music,
> and off-topic audio are removed, and it is redacted to our publication
> standard.

## Roster (in the room — 24; Luma check-ins ∪ live idea board joiners)

| Name | Business / role | Source |
|---|---|---|
| Austin Distel | Agentic Society (host) | host |
| Justin Day | SEO & agentic website systems | board winner + spoken [H] |
| Darby Rollins | Agentic Society (team) · GenAI University · comedy show | Luma check-in + board winner + spoken [H] |
| Oscar Hidalgo | Vindex Consulting · ops lead, Austin HVAC company | Luma check-in + board winner + spoken [H] |
| Sagun Sedai | Real Austin Report | board |
| Alan VanToai | AI Performance Lab | Luma check-in + board |
| Paul H. Graham | EG Capital | Luma check-in + board |
| Andrew Sponsler | AI & automations consultant | Luma check-in + board |
| Peter Hewlett | PCH Conversions | Luma check-in + board |
| Chrissy McDannell | The Magnolia Firm | Luma check-in + board |
| Ashley Kirchner | Decked Out Factory | Luma check-in + board + spoken [H] |
| Doug Greenberg | PNW Advisory · LocalFrequency.app | board |
| Alexander Felice | Threadquarters | Luma check-in + board |
| Mike Couvillion | Alexa Guru | Luma check-in + board |
| Jesse McElvain | Distru | Luma check-in + board |
| David Boskovic | Obvious | Luma check-in + board |
| Lucas Ferrer | Imagenyx | Luma check-in + board |
| Wilson Smith | Layer Zero Launches | Luma check-in |
| Brian Folsom | Inventory Sheets | Luma check-in |
| Damon Flowers | Modern Operators | Luma check-in |
| Chance Forman | Lillybug Home | Luma check-in |
| Mitchell Zoll | Zoll Firm | Luma check-in |
| Jo Feathers | Wellness creator | guest directory |
| Cristian Aguilar | Agentic Society (team) | Luma check-in |

---

## Transcript (attributed · timestamped · signal only)

### [00:11:00] Welcome & how this event works

**Austin Distel [H]:** Okay, squad, we're getting started. Welcome in. A lot of AI builders in the room, a lot of business owners as well. Everyone's welcome here.

Today's mastermind is not just about what me and my team will share on ways that we're using AI. This is going to be time for you to participate. There will be an option for you to share your screen: we have a Zoom link, and if your idea gets selected, you'll get to share what you're building.

A few notes. The whole room is being recorded; there are microphones in the ceiling, and this gives us really good audio. After this event there's a triangulation of transcription between devices, and this allows us to create a microsite for this event. You can connect with your peers in the room after, on LinkedIn. Every tool mentioned gets backlinked. It will give you playbooks about how each build was made, and copy-and-paste prompts that you can give to your agent to set it up too.

Collectively, as a mastermind, steel sharpens steel. Share as much as you want. Those that participate at the highest level: we do a little award ceremony at the end. You get a trophy, you get promoted on our social channels, and emailed to our list about what you built, how you built it, with links to you. If you wouldn't like something shared, that's okay — just tell us to redact it. This transcription goes through a redaction skill that pulls out anything that feels sensitive, and you can always tell me directly too.

The mastermind has been running for the last 90 days; this is the fourth event. Everybody here was probably referred by a friend, and that's how we've kept the quality of the room high.

### [00:11:49] Fridays are for the future

**Austin Distel [H]:** We have a philosophy that Fridays are for the future. We are all so busy working IN our business Monday through Thursday, maybe Saturdays too. Friday is the day that we work ON our business and figure out what the frontier looks like. That's why you're here today: to figure out the latest things happening in AI and how business owners in Austin are using it to actually grow their businesses.

We're all inundated with reels and TikToks and YouTube and X and groups — so much information. This is that filter. This is to find the signal through the noise and to really figure out what's working, and not just what's cool and shiny.

**Austin Distel [H]:** My background: I've been in the AI space six years now. In 2020, my team got access to GPT-2 through our Y Combinator connections. We had been training teams on how to do great marketing — this was our third software company. That AI could autocomplete a sentence, and we went to market with one idea: we trained an AI on direct-response copywriting. That became a company called Jasper. We named it Jarvis at first, and renamed four months in. Four months in we already had a million-dollar run rate; by the end of the year we closed $50 million in cash collected. I was part of the first chapter, and left after the Series A — Jasper is now run by rockstar executives and supports the enterprise.

After exiting, I bought a gym and spa here in Texas — one in Austin, one in Dallas — and that's my playground for applying AI in a normal business. I've automated it to about one day a week of my involvement. The rest of my time is building this community of AI builders and business owners. We're 90 days in, with 25 paying members. My goal is to inspire the community by letting you be the poster child of what's happening in AI: real stories, real outcomes — revenue growth, expenses reduced, time saved. Outcomes, not activities.

**Austin Distel [H]:** I'll know this was a valuable use of your Friday if three things happen. One: you walk away with one meaningful connection — write that person's name down. Two: one brilliant idea — one signal through the noise that made this room worth the filter. Three: one decisive action — on Monday, what are you going to install, explore, or refactor about the way AI works in your business?

### [00:19:47] The three levels of AI

**Austin Distel [H]:** There are three levels to building your skill sets with AI.

Level one is where most of the world is: using chat as a conversational tool. A consultant, a therapist, rewrite my email, review this contract. It's only there when you go to it — a manual use of AI. 99% of people are here, and if you are, that's okay. My goal is to level you up as fast as possible.

Level two means you go from solo player to multiplayer mode. It's no longer working just for you — it's working across your organization. You have a company brain set up. AI is the fabric of the way your business works. Meetings are transcribed, maybe a Slack agent is listening in the background and updating project boards. You're using skills, artifacts, connectors. You're reviewing tools based on: does it have an MCP yet?

Level three: you have a fleet of AI agents — AI employees with a defined job in your business. They don't just work when you go to them; they operate 24/7, 365, on cron jobs or whenever you need them. This is the shared language of harnesses like Hermes, and multi-model orchestration. I call this role the agentic CEO.

### [00:26:13] The Agentic CEO

**Austin Distel [H]:** In the last nine months AI has become agentic: it can go out and do work, not just hold a conversation. It has access to your tools and intelligence, and it can control your computer — everything your mouse and keyboard can do, probably better than you can.

Most of us in this room are non-technical business owners, so our shared language is who, not how. You don't need to know how to do everything in your business. If you're always the one doing, you're the bottleneck. You need to create the identity of your agent, give it access to the things it needs to perform on your behalf, and review its work.

Software is now headcount. It's no longer just that we're paying salaries — we're paying for tokens. It's not that we're not hiring; it's that we're hiring the right people who use AI agents in their work and think like a boss more than an individual contributor: set up a system, automate, delegate, and move on to the next system.

Think about it through the lens of hiring: job posts, KPIs, tech stack, onboarding, training, one-on-ones, check-ins. On average it takes a senior hire about four months to perform at the level you hired them for. An agent, with a company brain and skills set up the right way, can be there within seven days or less.

Empower your team with positivity, not fear. The marketplace right now is fearful about jobs — sometimes that's true. But give your team an honest shot: teach them the tools, level them from level one to level two, give them the company brain and agents for their department. As the CEO of your business — or the AI builder your company assigned to figure this out — the agentic CEO is the role that defines how the company runs on AI.

### [00:30:27] The company brain

**Austin Distel [H]:** The milestones work with the levels. Level one, you use the tools, but in a silo they're only as smart as the LLM. Level two is where you build the company brain. We're recommending Google Shared Drive — not Obsidian, because that's often solo player. Google Shared Teams Drive: governance, security, permissions, scalability. If you already know how to invite somebody to a folder and set view or edit permissions, you already know how to run it, and it works no matter the technology stack — Claude or Codex, it doesn't matter.

What matters is that you own your data. You're not renting a platform; you own your company intelligence. If you build the right company brain, your company is more defensible, sellable, scalable — you're no longer the bottleneck of information. When you onboard employees or sell the company, your Drive is the container, and you point LLMs at it. Whenever the tides change of who's winning between the labs, you still win. The third stage: with the foundation built, you bring in advanced AI agents like Hermes and point them at it.

*[Community overview compressed: training in the community classroom (a seven-day Claude crash course), the Agentic OS — plug in one MCP and a recruiter agent named Harper installs pre-built AI-employee skills — weekly Tuesday calls, twice-monthly member boardrooms, and an October 15–16 hackathon in this room. Membership enrollment details from the room are omitted here; ask Austin for current programs.]*

### [00:40:00 ~] Member stories from the room

**Austin Distel [H]:** Justin has made his SEO agency fully agent-driven — from 30 employees down to nine, with AI agents running everything from sales-call review to client success, taking on more volume than ever, and a lot of margin back to the founder.

Erin has a functional dietary clinic with 12 dietitians. One big win was getting her COO an agentic dashboard to get ahead of client churn: renewals surfaced, contact points tracked, nothing slipping through the cracks — more revenue retention in 90 days.

Rohan interviews CEOs and creates content for them on LinkedIn — he clips long interviews into viral moments using AI.

Isaiah is a developer with 15 to 20 autonomous agents working for him. This community has been his way to get away from the screen: "I've been building solo for 90 days, working early mornings and late nights talking to robots." Being together is better.

Prashant is a real estate developer landing wholesaling deals using AI SMS agents — appointments set on his calendar from a list, with pipeline management next.

Oscar landed $400,000 in revenue in the last 90 days with an AI outbound prospecting agent reaching out to building managers. He won our award last event for it.

Roman, my fractional COO — eight years with Disney+ — created a fleet of AI agents in Slack that revolutionized our project management: transcripts read, tasks assigned, priorities ranked, follow-ups in Slack for every person on the team.

### [00:52:00 ~] The live idea board — submit and vote

**Austin Distel [H]:** Last mastermind I brought out a giant whiteboard; you'll notice it's not here this time. Get out your phones and scan this QR code. This is where you submit the AI ideas you've gotten results with in the last 90 days. Real outcomes, not "I'm thinking about doing this thing." Then we all vote — you get five votes each. There's 19 people in the room and we have 14 ideas on the board [more arrived before lock — 16 final]. If your idea gets voted, you get speaking time, we highlight you on social, and your story goes on the leaderboard.

Look at this board: 16 outcomes where AI agents literally improved a business in this room. A lot of the people on this board are non-technical. What this shows is that we are resourceful enough to solve problems in our business using AI. And the voting shows what is the most in-demand need — demand and supply. Justin, Darby, and then Oscar will share what they built and what you voted on. This is not just a presentation — it's a conversation. Pressure-test it. "Why did you choose that versus this? I do it this way. Is your way better?"

### [1:01:33] Deep dive 1 — Justin Day: the self-growing website swarm

**Austin Distel [H]:** Justin, you're always on the leaderboard, bro. You built an agentic website system that learns from its own data, grows traffic while you sleep, and self-improves to increase conversions. Take it away.

**Justin Day [H]:** Going into the problem: for the last decade people have been building websites on WordPress and page builders — built for what looked good, not for what converts. My background is a global SEO and website business; over the last ten years we've generated a lot of results for clients, and what we found is patterns. We know what works from a website conversion standpoint. If you have the right setup and the right steps, you get the right website. Otherwise it might look cool, but it's most likely not going to convert.

So one: a build system that goes from a transcript of an onboarding call to an output that's basically pixel-perfect and built to convert. Everything I'm talking about today will be available in the community in a couple of months — you can have my entire system.

But that's still a static website. A website built with AI should self-learn over time and self-heal. There's no reason, with all the APIs we have, not to. Growing traffic — blogs, content, keywords, what the market needs — that's the data. We plug that data in and make intelligent decisions about what pages to push and what topics are drawing traffic. A visual heat map shows where people stop on a page and bounce, so I know where to update copy or a button. And leveraging Sentry as an observer to see when things break: self-heal. No reason to have a plugin or a person come fix something when you've got AI. I built a swarm system to do that: self-growing, self-CRO, self-healing.

**Justin Day [H]:** The build system stages: an onboarding call captures who your target audience is, the brand guidelines, who's NOT the right fit, design examples. That becomes the North Star document — everything about the business that we always root back into. Then the page playbook: are these the right pages? What's the site cap, so I'm not building random pages? We're here to make money and drive revenue, not to have pages that don't do anything. Then architecture: every heading, every button, every image, per page — home, about, contact, FAQ each have their own best practices. Then copy, with SEO implemented at the same stage — keywords, headers — so by design time the technical and the SEO are dialed in. Then design from a component library, according to the client's North Star. Then A/B variations of the first three core pages — home, about, service — before developing the rest. That's the process that gets you a winning website.

I also built a pixel-perfect cloning project: if you're afraid to move an existing site, enter a URL, clone it over, and improve it from there.

**Justin Day [H]:** What a self-learning system looks like is a loop: data sources feeding processors, feeding an intelligence layer, feeding an action layer, deliveries, and a feedback loop. This is what our second brain looks like as a company, and the website gets a mini version of the same brain. The action layer is: what do I need to do today that's going to drive revenue for the business?

**Attendee [M]:** How do you know the conversion rate is actually converting? You could be self-healing in the wrong direction.

**Justin Day [H]:** Quality gates. It's not fully autonomous — you want a human in the loop, so it sends a Slack message to approve things. Everything is tied into the CRM, so we're going on correct revenue data, not just traffic. We're not going to touch anything based on traffic alone. And KPIs per page: you need, say, a thousand people landing on a page with 30 to 60 seconds average time before you make a move. If you're optimizing off ten visitors, you're not optimizing.

**Austin Distel [H]:** Justin, what does this replace?

**Justin Day [H]:** An entire team. From an autonomous website standpoint it replaces growing your traffic organically on Google and AI, and the website developer — the days of paying someone $50,000 or $100,000 for a website build are gone. An agency running the 2015 playbook is completely dead. And it replaces a lot of subscriptions, honestly — you were feeding Google Search Console and Analytics raw data into a keyword tool just to patch insight together.

**Austin Distel [H]:** And who is this not for?

**Justin Day [H]:** That's a good question. If you're an influencer or someone who's more branded, I wouldn't necessarily implement this. It's mainly for local service businesses — and enterprise, for the data insight over time. The biggest ROI is local service.

**Austin Distel [H]:** The brand is going to care a lot about how it's displayed — but a local service provider is like, all I care about is my phone is ringing, leads are coming in, and I'll do whatever it takes to win on search. And for a brand like a Nike.com, the brand manager and website owner could still use something like this to inform their product board or feature implementation.

**Justin Day [H]:** From a data-aggregation standpoint it applies to everybody — the amount of insight you get on the back end, you'd never assemble by hand. The self-healing CRO part is more in line with local service.

**Austin Distel [H]:** Let's talk about the tech side. Most people's website is on WordPress or Webflow. What does this look like — hosting, deployment, where's the repo?

**Justin Day [H]:** Everything I build is using Claude. We push everything into Cloudflare — they've got a scalable model from an agency perspective, and Workers lets us deploy agents onto the website itself, versus static hosting. Sentry for oversight — you see Slack messages on the fix; that's the self-healing mechanism. PostHog is our self-CRO: it has an MCP and an API, it's the API-friendly alternative to Hotjar, and it's about $30 a month rather than a couple thousand a year. Plug in Analytics, Search Console, Tag Manager as data points. Repo on GitHub and push from there.

**Attendee [M]:** When you want to change the website, what's the workflow — admin console, Cloudflare, or talking to Claude?

**Justin Day [H]:** Internally we just update it by talking to Claude, which is so freaking smart. For clients who want to edit their own sites, I built a client-facing chat interface — the visual chat-editing experience, versus the 2015 pencil-and-drag editor — with token-based pricing, so an edit costs cents.

**Attendee [M]:** Does this work for dentists, physical therapists, medical specialists?

**Justin Day [H]:** Absolutely — we're working with a lot of doctors. And the shift in AI visibility is massive. Any business showing up in results matters, whether it's AI or classic search.

**Attendee [M]:** What about landing pages and funnels?

**Justin Day [H]:** The methodology lives there — I run PostHog and Sentry on my own funnels — but I'd be cautious about letting paid-media pages self-improve. That's paid traffic; you've got a lot more to lose than organic. I want reports there, and I decide.

**Attendee [M]:** What surprised you most?

**Justin Day [H]:** Honestly, the hesitation from clients — local service owners who've had their website forever and believe in it. So we meet them where they are: keep the investment, clone it over, make it better over time. But from a results standpoint, it's like just putting gasoline on a fire. You can have a website that's been doing no traffic and then just blow up over 90 days, because you've got all the right signals.

And here's what I'll really dig into at my talk next week: search is moving to agents. Since June, agents are going out to websites to book services on people's behalf and bringing data back into the results page. In the next year people increasingly won't go to websites — they'll stay inside the results, with custom dashboards built in real time, or an agent booking and calling for them. If your website isn't agentic-ready, you're already behind. And to that exact point: we are no longer designing for humans — we are designing for bots across the board on the web. That's Web 4.0. It's just agentic-ready.

### [1:27:59] Deep dive 2 — Darby Rollins: the content production & distribution agent

**Austin Distel [H]:** Darby, you're up. Ten people voted for your AI content production and distribution agent: it helps edit social clips, create thumbnails, and posts to social automatically across platforms for consistent daily content, with variations of short-form intros to post and test.

**Darby Rollins [H]:** The agent I finished recently is about taking content back from production and creating all the variations we post. Context: I run a comedy show based here in Austin. We've done shows in other cities, and we're building toward a big 2027. The problem with that business is capturing content professionally — which takes a lot of time and money — and then: how do we consistently get short-form out there so people see what the show's about and come to the next one?

The last couple of months we took four filmed shows and worked with a professional editor to create the long-form and a bunch of short-form clips from it. Once I get a batch of four to eight clips per episode, I download them to my machine, and I have an entire video-analysis agent built with Claude and Codex — using Hyperframes from Hugging Face, plus libraries Claude researched for me — loaded with our brand style guide: who we are, what the show's about, who we want to reach, our colors, the vibe. It scans the whole episode, creates tons of thumbnail variations, clips and highlights both vertical and horizontal, and builds the front matter for each so we can keep clipping over time and test variance across platforms.

The other thing that kept me from posting consistently: I don't want to be on social media every day, in every platform, posting manually. So once I'm ready to batch two weeks of content, I use Blotato — an MCP and API integration for posting. It can post to nine platforms — Facebook, Instagram, TikTok, Threads, Bluesky, X, YouTube, LinkedIn — and I post to about six right now. I've got multiple weeks batched on all these channels, some we've never been active on, and people are responding and subscribing. Our next show is in October, so my goal is a base of organic data: which videos do people click into? What's consistent in those titles? How long do they watch, on which platforms? The Blotato MCP gives you that data back if you post through it.

Blotato also just added a ManyChat-style feature: on a schedule it scans your posts, and when someone comments it replies and sends them your link. The week before an event, the last thing I want is to live on my phone sending ticket links. And I hardly ever log into the actual dashboard — it's all MCP-driven. I just ask Claude: give me an update from yesterday; who responded; reply to everyone who commented with interest.

**Austin Distel [H]:** How many of y'all do short-form content? Okay — very applicable. What part of the process does this NOT touch?

**Darby Rollins [H]:** I don't think you should be outsourcing the video edit itself to AI at this point. The taste for what's a good hook, a good intro — I've tried it, and it's hit or miss. We went the route of a professional editor who does YouTube content, with focused feedback. Once that content's done, I want to be able to scan it and get little clips from it, so I can repurpose it and get a lot more bites out of the apple: one short across six to nine platforms, plus three-second intro variations to test, and 60 days of data to tell us which hooks hit — which then informs the next shows, the camera angles, how we frame questions. Next year, if we're doing a show every month, I'm not going to want to be in the thick of development mid-season. That's why I spent the summer building this.

**Attendee [M]:** Two questions real quick — HeyGen and the tools in this space, and what does the stack cost?

**Darby Rollins [H]:** I don't actually have a HeyGen subscription — when they introduced Hyperframes, they just said here's the thing, download it and use it. I don't think I've ever paid them anything for that. Blotato, I'm on a founding member plan at about $20 a month; their lowest is around $29 — still way too cheap. And Claude and Codex — I'm on the $200 plans for both, but I use those across a ton of projects.

**Attendee [M]:** I had editors doing short form, but they charged a lot for one short out of a long form — so I fired them and use Riverside; its magic clips get you close, then you tune a little. But I can't get AI to build a good cold open, even feeding it the best transcript moments.

**Darby Rollins [H]:** We just decided not to use AI for the edit. But it's getting better at finding moments — I ask it to find shots where the audience is reacting or making an expressive movement, and it pulls still shots I'd have spent hours scanning for. For our February show I did that by hand; I'd rather not again.

**Sagun Sedai [H]:** With HeyGen, some people make an avatar that looks and acts exactly like them. Instagram is pushing authentic content and flags AI — does that affect performance?

**Darby Rollins [H]:** None of ours is generated — it's all humans in every frame; the AI only analyzes and edits. I made an animation of my robot mascot with an ElevenLabs voice, comedian pacing — really cool, but I'm not going to have that replace me. I've decided not to make myself an avatar: there's a lot of trust that can be broken quietly taking that approach. For a comedy show, we highlight the human side.

**Attendee [M]:** My wife tried the avatar route — no measurable impact, but there was still an uncanny valley; you could sense it wasn't quite right. When she switched to a clearly cartoon avatar, people understood, and it was fine. One thing that did work: re-syncing a good take over bad audio — still her voice, still her face, just synced up.

**Justin Day [H]:** If you're leading with trying to mislead people and make it look like a person, it won't work. If you're leveraging it as a funny thing, great. But if you try and trick people, it ain't gonna work. People know.

**Sagun Sedai [H]:** Video editing is this final boss — a really challenging frontier for AI because of how subjective it is. It's not quantifiable; you can't do pass/fail evals on what a good edit is, the way you can with code. Every week there's a new tool on X promising to solve it. I use Opus Clip for my podcast and it's almost spot on — you just trim some angles — as close to usable as anything.

**Darby Rollins [H]:** And for cleanup, Descript — bad gaps, studio sound. Even this week, my editing agent analyzed a workshop recording and said: I found Austin sharing his open Canva link on screen; per your protocol I've blurred it out. It found sensitive information and redacted it on its own.

### [1:49:11] Deep dive 3 — Oscar Hidalgo: ten Hermes agents running the back office

**Austin Distel [H]:** Our last winner is Oscar. He built a team of ten Hermes agents that replaced $200,000 worth of back-office payroll for the HVAC company he runs here in Austin. Background: Oscar has now joined the Agentic Society team as our Hermes agent builder — he's done such a great job defining business use cases on Hermes that he's helping with client installs and creating courses on it.

**Oscar Hidalgo [H]:** A brief intro: I'm not as technical as I need — even a lot of the stuff Justin's doing, I'm like, dude. I do operations. I've been running businesses for about ten years, and my day job is managing a commercial HVAC company — long three-to-four-month projects with a lot of operational management.

The big thing for me: once I found Hermes — for me, the replacement to the harness I kept having to fix every day — it's really good at determining a role. You define what an actual need in your business is and turn it into its own agent, so you're not using one chat for every single thing.

This is my Hermes setup. I use the desktop app, but my actual Hermes lives on a remote server — always on, 24/7, running whatever automations I want. I access it from Telegram or Slack; that's why I don't carry the laptop. They recently released remote gateways, so I get the best of both worlds: the nice desktop UI for deep work, Telegram on the go.

Each of my agents has what's called a soul — a predefined set of instructions for how that agent performs. That's what puts it above a plain chat project: you name them and treat them as actual AI employees. My default agent is Eve. The first rule of protocol for any task: open the company brain and read, in this specific order, the agent instructions, the soul, the company brain — then start the task. I'm predefining the context so the agent never has to reconfirm who I am, what my company does, what my branding is.

My second brain syncs between my MacBook and the virtual machine using Syncthing, so whatever changes I make locally reflect on the server — all my agents work from the same context. There's a roster: every agent Eve can route to, and what their abilities are. A rebate-project task goes to Rex; financial formatting goes to Clara. There's a routing rule that assigns one specialist per task. And there are overall walls. One of my favorites: no em dashes on anything external — emails, blog posts, anything a customer sees — set once, for all agents.

**Oscar Hidalgo [H]:** Why sub-agents? One agent that does one job really well, just like an employee: its own SOPs and its own set of tools. If I ask for a payroll report, I don't want it opening Chrome and taking screenshots trying to log into my payroll software — I want it using the MCP I connected. That's why it's not one generic agent.

His CRO agent for the HVAC company: weekly content batches (scans the photo folder, unifies reviews, upcoming holidays, branded previews), watches the Google Business profile, works the SEO content backlog into blog posts on the website, prepares marketing calendars.

Rex is the one I use most — multifamily projects. We do apartment complexes; the current one is 448 units, and on every unit we capture before-and-afters, pressures, superheat and subcool readings — things I don't fully understand, but the city requires a spreadsheet at the end of the project with all of it organized. This was one of my biggest operational nightmares; we used to have an operations assistant helping with it. Now: the field team submits the same Connecteam forms they always have — photos, readings, step by step. Rex grabs every submitted form and updates our Airtable, so I can see at a glance which units were completed when, with the photos preloaded — and the data is accessible to every other agent, because Airtable is flexible for agents.

That's the pattern: I've translated most of our applications to agent-first applications. Airtable instead of Trello. Superhuman for email, so every agent uses the same MCP. Workable for recruiting — an amazing MCP — so a recruiter agent can send onboarding packages and 1099 contracts. The more you do this, the more they become actual employees: they use the tools directly through MCPs instead of relying on a UI and clicking on things.

**Austin Distel [H]:** Oscar — you were saying this in the boardroom earlier. How many of your human employees know about your AI employees?

**Oscar Hidalgo [H]:** Zero. Just the owner and myself are the only ones that know about all the agents.

**Austin Distel [H]:** So even though he's built an agentic company, the HVAC specialists don't have to use AI. They fill out a form, and the admin work behind it is done by AI. How do you think about orchestrating teams of AI agents AND humans?

**Oscar Hidalgo [H]:** It depends on your industry. HVAC technicians — even some sales roles — struggle to even update a CRM. They don't want to log into a computer; they want to be on the phone, doing stuff with their hands. So understand what you can automate without adding friction that creates mistakes. As the agentic CEO, it's your new job to decide at what point this collaboration lives in the background versus involving the actual team with a dashboard. For us: the field team's workflow doesn't change at all.

**Attendee [M]:** What's your context methodology for the brain?

**Oscar Hidalgo [H]:** I already had a second brain before joining, then applied the one we teach now and mixed both. It's a folder structure: an AI-team folder where each agent has its own folder with its soul — all drafted by Claude proposing "here's how I think this agent should operate," and me editing. A company folder with decisions: at the end of every session, agents record notes AND decisions, so they can refer to decisions later. Daily notes. An outputs folder for files we've made. A people folder — so "prep me for a meeting with [the owner]" just works. Start by brainstorming the structure with Claude; you can go back and forth.

**Attendee [M]:** Why a VPS and not a Mac Mini?

**Oscar Hidalgo [H]:** A Mac Mini is probably one of the best solutions — it adds iMessage, Apple Reminders, local files. I just already have a powerful MacBook, and the VPS is about eight dollars a month — it works just as well because all the agents' work runs through APIs. Cloud providers have backups on the backup, so my agents basically never go down — which matters once you're running cron jobs at night. You don't want "my machine wasn't plugged in and now my business is struggling."

**Sagun Sedai [H]:** Do your agents run on a single server with different containers — or eight different containers or servers? And are they eight profiles within one Hermes? Can they run concurrently?

**Oscar Hidalgo [H]:** Single server, one container, and Hermes calls them profiles — each with its own soul, operating differently, and yes, they run concurrently. Each can have its own gateway: a separate Telegram channel or Slack bot per agent. Each profile can also run a different model — an agent that needs research on X can run one model; an agent making images can run another. My HVAC company pays for the agents related to it, and my consulting company pays for the others — one harness, split by tenant.

**Attendee [M]:** Are the agents talking to each other?

**Oscar Hidalgo [H]:** Not directly yet — most of the work routes through the orchestrator. There's an interesting project from a couple weeks ago — Buzz, by Jack Dorsey — like a Slack for agents. I've been playing with it, but nothing direct in production.

**Sagun Sedai [H]:** With this configuration, when you want to connect a new tool, is it a single connector configuration? And your API keys — one .env across all of them?

**Oscar Hidalgo [H]:** Connecting tools to Hermes is open source and a bit rudimentary — recently I've been recommending Composio.dev (or its competitors) to connect all your MCPs in one place and hand them to your whole fleet. And yes, one config: I just ask Eve to assign tools to whichever agent needs them — "Rex is going to need this tool, add it to his config." I'm using GPT as the brain — the $100 ChatGPT subscription runs all of this. And each profile can run a different model: an agent that does X research can run Grok; an agent that makes images can run Gemini. My HVAC company pays for its agents and my consulting company pays for the others — one harness, split by tenant.

**Attendee [M]:** How do you design a new workflow?

**Oscar Hidalgo [H]:** I start with mapping it myself — whiteboard, paper — the same thing I would have done if I was hiring an employee: where is it actually needed to make this its own person? Then I brainstorm it back and forth with Claude. Once it's fully defined: Claude can access my VPS through Tailscale, so I just tell it, hey, build it on my VPS. One trick: when you download Hermes, it comes with skills specific to how Hermes works. I made Claude read all those skills and turn them into Claude skills — so now Claude is a Hermes expert builder.

And boundaries, at the end of every soul: do not execute specialist work when a specialist owner exists (that one is for the routing agent). External communications: drafts only. Never send a meeting invite without my approval. Never move money, pay, sign, file taxes, or commit financially. The owner's confidential information stays in context only.

**Austin Distel [H]:** Give Oscar a round of applause. Oscar, what courses are you creating in the community?

**Oscar Hidalgo [H]:** Getting started with Hermes — running it local or on a virtual machine, what brain to use, getting your first agent going. Then specific agent profiles: the SEO agent, and the outreach agent for commercial properties — it used Apollo to find the right email contact information and send a cold email, and we got a good response from that. On agent email: I started with drafts from my own inbox; a tool that has been very helpful is AgentMail — now the default way to give agents their own email address, on their own domain. And for cold email, always a separate domain.

### [2:14:16] Awards, insights & gratitude

**Austin Distel [H]:** Here's the fun part. Those that shared and got the most votes — give it up for Oscar! Before we hang out in the lobby, I want to take a moment to share among the crew: what are some of the insights you wrote down today? Keep it to fifteen seconds.

**Attendee [M]:** New to me: in Hermes you can spin up agents without creating containers or servers — they're just a profile. And shared connectors, so you don't keep reconfiguring tool connections for every new agent.

**Attendee [M]:** My takeaway is the Google Shared Drive for the company brain. I'd heard pros and cons, but it sounds like you're sold on it.

**Austin Distel [H]:** It's something we all mostly use — you don't have to re-solve security and privacy, because your business is already set up on it. And you can assign AI agents as new email-address employees and give them access to only their folders.

**Attendee [M]:** Are your documents markdown files now?

**Austin Distel [H]:** Both. When a rich PDF or spreadsheet lands in what we call the mailman folder, it gets read once, documented in an MD file — like a readme — and filed. When an agent enters that folder later, it reads the markdown like a table of contents instead of ingesting everything. That saves a lot of tokens.

**Ashley Kirchner [H]:** I'm building my headless Shopify stores right now, so PostHog — the heatmap piece — is the one I'm taking. I have this idea where the site actually changes for people based on their preferences and where they're located, and learns from that. It's started, but there are back-end tools I've been missing.

**Justin Day [H]:** Shopify out of the gate has a really good agentic platform — on AI readiness it's super solid, and it already integrates directly with ChatGPT so people can shop from there; Google does the same.

**Attendee [M]:** Side comment: agentic ChatGPT plus Shopify turned on automatically for me this year, and it made money — not a lot, and I didn't touch it. I'm watching those tools just produce income.

**Justin Day [H]:** Leverage the Google Analytics MCP to review your setup — make sure attribution runs down to the dollar. And Google Search Console just launched an AI tab under performance showing which content is driving AI traffic — plus, this week, how people search for you on Instagram and YouTube. Social signals are now a huge part of SEO.

**Austin Distel [H]:** Who else impacted you today? Who's in that middle column — the meaningful connection? Shout it out; give gratitude.

**Attendee [M]:** Gratitude to Justin — I got a couple ideas I can apply in my short-form content work.

**Sagun Sedai [H]:** Thank you to Austin for the invite.

**Darby Rollins [H]:** I want to shout out Oscar for reminding me about Composio. I forgot how easy that is — super useful.

**Austin Distel [H]:** Amazing — that wraps up the mastermind. We host this the first Friday of every month, and first-time non-members can come back every ninety days. Members, keep coming to Tuesday mornings. We'll be in the lobby for the next thirty minutes — please eat another taco.

— End of transcript —