# AI Boardroom — Agentic Society Members · Clean Transcript
**Date:** 2026-07-24 · **Type:** AI Boardroom
**Location:** Industrious East 7th, 924 E 7th St (Conference Room A), Austin, TX
**Duration captured:** ~3h 26m (in-room, one shared mic + a few individually captured voices)

> **How this was built:** the room was captured on a single in-room mic, so "who
> said what" was reconstructed from self-introductions, direct address by name,
> and business-description matching against the member roster. A couple of
> voices in this session could not be tied to a name with confidence — where
> that happened, this transcript describes the contribution without guessing
> who made it. Quotes are lightly cleaned for readability; timestamps are from
> the raw Zoom recording timeline.

## Who was in the room
| Name | Business | Notes |
|---|---|---|
| Austin Distel | The Agentic Society (host) · The Tox | facilitator |
| Darby Rollins | Gen AI University · Agentic OS | co-taught the Company Brain session |
| Ravi Parikh | RoverPass · agentic accounting firm · Maker Square | new member |
| Justin Day | Day by Day Productions | |
| Jay Douglas | KeyFive | new member |
| Mike Couvillion | Alexa Guru · AI consulting for PE-acquired software companies | new member |
| Prashant Vanka | Real estate development | |
| AZ Moyer | AZ Does Taxes | |
| Will Sayasenh | Video & Photography (Agentic Society team) | |
| Rohan Karunakaran | Frontier Studio | |
| Cristian Aguilar | Agentic Society | |
| several other members and new guests | — | present but not individually featured below |

---

## Transcript

### [00:02:58] Before the room opened — the format idea (Austin & Darby)
Before the session started, Austin floated the structural change to Darby that
he'd announce to the room later: "What I hear a lot from people is they want to
use this time to connect with people in the room. They're not always coming
here just to get a presentation — it's 'let me share my work, collaborate with
you and your work.' The Academy has its own place for teaching stuff. We're
using this room as a way to do a little more personalized, hands-on work." The
shape he described: teaching in the first half of the day, then the second half
of Fridays open "where everybody can come work here together, as deep work
together." He added: "Our online members are not going to get very much
hand-holding from me and the team. But us being here in the room with you guys
— that's more of that hand-holding touch. There have already been so many
partnerships and deals that have happened, but people have to get to know each
other. So I'm really focusing the container on connection, community, and
personalization."

### [00:33:17] Format & welcome — Austin Distel
Austin opened the room by placing the Boardroom inside the larger system:
"Every container in Agentic Society has different purposes. Tuesday calls — AI
Breakthrough — that's for you to try to get your problem solved by the team.
Friday boardrooms are for our Premier members, like you guys, who want closer
access, more involvement, more connection, more deal flow. That's why you're
here — there are so many online communities out there, and you chose to be
physically in the room. Everybody here is a really sharp business owner, so
you guys have a lot to connect over."

"My intention is for you guys to build connection in this container. Afterward,
y'all are welcome to use the whole co-working space for the rest of the day —
if you want to continue your conversations, or do deep work together, or
collaborate and co-work, out there is a great time to do that. We'll end at
12:30."

On the teaching half: "Part of this is that you get early access to all the
training we do inside the Academy. What gets published abroad gets first built
here — and personalized with you all. We help you with the installs. It's that
level of hand-holding access that people going through our online content
won't ever get."

Then the day's agenda: "Today's focus is building your company brain. This has
been a hot topic for 45 to 60 days now, and Darby and I have worked a lot on
creating a very simple version of it. There's a member in Agentic Society that
has a really awesome software company that's about building company brains for
large companies. The way that we're building this one is specifically to be on
free tools — Google Drive — in a way that AI agents can read really well, and
in a way that your team has access to it. Because if you're the only one
contributing to your company brain, you're doing your whole company a
disservice."

### [00:36:36] New member intro — Ravi Parikh
"I'm Ravi Parikh. I have three companies. The biggest of them is called
RoverPass — we're a management software company for the outdoor hospitality
industry: RV parks, glamping resorts, campgrounds. We have about 10% of all
U.S. campgrounds using us. Then I have an agentic accounting firm — we do tax,
bookkeeping, accounts payable, and all the boring stuff that small business
owners don't want to do. And I recently relaunched a training company I
started back in 2012 called Maker Square, which was a 13-week software
engineering bootcamp here in Austin. We sold it in 2015, and I relaunched it
to teach people AI — mostly non-technical people."

On why he joined: "I feel like I'm either on X.com learning about AI, or just
trying to figure it out myself, and I don't really have a peer group that is
doing the same thing as me. Sometimes I feel like I'm going crazy. Hopefully
you guys are using AI every day like me."

### [00:38:26] New member intro — Mike Couvillion
"I'm CEO of my own consulting company — and then, in a completely tangent
thing, I'm also co-owner and COO of an interior design company. I've been a
CTO for 30-plus years at a variety of SaaS companies. We've been fortunate
enough to have multiple exits — hopefully a fifth coming up soon. I started
life as a programmer and systems engineer, like a lot of guys, moved up the
chain, and semi-retired a couple of years ago after the last exit. I'd been in
expert systems and machine learning systems in several industries, and when
the whole LLM and AI thing started, I dove into it really deep."

His consulting practice: "I consult with private equity firms, mostly those
that acquire software companies — helping them find CTOs, and showing them how
to actually roll out AI into your company and not blow 95% of your budget in
the first month. Most of the time, a new CEO comes in, or the private equity
company comes in and says: for the number of people I'm paying to write code,
we're not getting the velocity. I've never heard any CEO say, 'yeah, my team's
too fast.' And it's usually going in and finding out that 85% of the time,
it's not the software team that is the problem — they are the symptom. Your
CEO and your COO and your CFO and your CTO all have a different idea of what's
number one in the company. Help them figure that part out, and then you can
identify, okay, here's the 10% of your engineering team that needs help —
let's go fix that."

"In the last six months it has been almost exclusively: 'I bought a bunch of
Claude licenses, we're spending a lot of money on tokens, but nothing's
changed. No deadlines have come in, nothing's getting delivered. Why?' To me
it's just like the early days of AWS, when you gave developers the keys and
the next thing you know you've got a runaway EC2 bill, because everybody
fires up a server and nobody shuts one down. It's the same if you don't have
that roadmap, that harness, that strategy. AI is going to make you faster
even if you're doing it wrong — it'll just make you do it wrong faster." He's
also built two enterprise platforms since diving in, with about a hundred
customers on them now — "paying for my Claude license."

On the interior design side: "It's amazing how much building a house is not
that much different than building software, except for the final deliverable.
A lot of the problems are the same. Like a lot of traditional business owners,
she's leveraging AI in the obvious places in interior design — concepts, art,
drafting — but it's really the business operations where I work: how do we
take all those spreadsheets of stuff and make that into a workflow that,
instead of taking three hours to crunch, takes half an hour? Now my designer
can go and do what we actually get paid for." His wife runs a high-end
interior design firm for custom homes; Mike runs its operations. "It's
interesting, because it was not the world I came from."

How he found the Society: "We work with Chance, actually — Chance is who I
talked to. He and I were at a company warming picnic, and somehow the topic of
AI came up, and the next thing I know Chance and I are holding a TED talk for
all these business owners. And Chance was like, 'there's a group you need to
join.'"

### [00:42:51] New member intro — Jay Douglas
Jay Douglas, founder of KeyFive: "The reason I joined the Agentic Society is
that this is a new frontier of technology. It reminds me of the early days of
the internet, where patterns and practices were established in rooms like
this — there was no communal right way to do anything. That's why I'm here: to
learn, and contribute when necessary."

### [00:43:18] Two-week wins — how it works — Austin Distel
"We run these boardrooms in two-week sprints. So between this week and two
weeks ago — what have you done with AI? What have you accomplished? If you
have a win you'd like to share, we do popcorn — it's not going around,
everybody doesn't need to share — but if you got a win and you want to be
celebrated in that, we'd love to celebrate you. If you had a breakthrough, a
new channel, a new AI employee that you created, a new outcome, a client you
landed because you used AI agents — this is the time to share that, because we
want to know what's on the cutting edge of applications of using AI agents.
If you just saved yourself ten hours a week, time to share that. If you just
built a really cool dashboard that gave you an insight and you're like, wow,
this is shifting the direction of my company — time to share that."

"This is the filter through all of the noise that is out there on the internet
about AI — 'this killed that' and all of it. This is the room that's the
filter for: hey, I actually got real results using it."

### [00:44:44] Win — Ravi Parikh: the RFP-automation agent system
"I've got stuff," Ravi opened. "This week I've been working on a series of
agents that automate our RFP bidding. We bid on RFPs for counties and
municipalities, and the process of finding the RFPs, vetting them to see if
they match our software and all the technical requirements, then writing the
RFP — that used to take us two people maybe 80 hours each. These are
100-page documents you're writing. There's a lot of analysis that needs to be
done — security, SOC 2 Type 2 compliance, all these different bells and
whistles. And most of the time you don't have everything the RFP needs, so
there's some custom engineering that has to happen."

The pipeline, step by step: "We use a company called BidPrime that goes out
and sources RFPs — we put keywords into it and it pops up RFPs. But that's
actually the easy part. The hard part is analyzing the RFPs to determine
whether or not we should bid. We'd kind of built a company brain already, so
there's a series of agents: the first agent ingests the RFP data — the bid the
county or municipality has released. It looks at our codebase to determine
whether we actually have the requirements, and whatever requirements we
don't, it creates user stories for those, drafts all of the implementation —
everything we would need to do to build any feature gaps — then sizes how big
the feature gaps are, compares them against our committed product roadmap,
and gives us a bid-or-no-bid decision based off a scoring mechanism. That's a
couple of different skills, and Claude connected to our codebase and GitHub."

"From there, a couple of agents extract the format of the RFP, and it drafts
the entire RFP for me. Then there's an adversarial agent that goes in and
looks at the final draft: okay, did you overlook anything? Did you hallucinate
at all? It loops through that adversarial agent about ten times until it's
more or less ready for us to manually human-review. And the final step is a
design agent — we use a tool he named as Magnify to generate images, and it
takes those images and puts them into the RFP — and then we use Playwright:
there's an agent that figures out what screenshots of our software we need,
opens up our software, and just grabs the screenshots."

The result: "That whole process used to take us forever, and now it takes us
about 30 minutes. These are six-figure-a-year government contracts. You're
going to bid on twenty of them and maybe win one, right? So if you can bring
the cost to do it way down, the math is super in your favor. I spent most of
this week on it, pretty much deployed it, and submitted our first one
yesterday — to Miami-Dade County."

Asked whether the same workflow buys him delivery efficiency: "Yeah, because
it identifies the feature gaps. And we also replaced Jira — we built
essentially our own version of it. We have a company brain platform called
CENTCOM that centralizes everything we do: every department has a section —
finance, sales, support, engineering. The user stories pop in there and go
through already-established skills that refine each story down to a spec and
an implementation plan. When we do sales we commit certain dates to
customers, so it locks that feature into a quarter, and then it's playing
Tetris to figure out: can we actually feasibly deliver this based on our
current PR velocity? And the thing that's interesting is our capacity has
changed so much over the last six months that honestly we don't even really
know what our capacity is anymore. We used to ship 40 PRs in a quarter; this
last quarter we shipped 180. Maybe next quarter we'll ship double that —
because the models are getting better, or we're just learning how to do this
better too."

### [00:50:46] The Agentic-CEO mindset — Austin Distel
"I want to give recognition to Ravi for having the kind of mindset an agentic
CEO has. It's the goggles we put on in our business to spot constraints and
bottlenecks, and then we try to solve them with agents instead of solving
them with manual work or delegation. This is very high-level thinking: 'I see
there's so much opportunity to bid on huge government contracts — the
opportunity size is huge — but I know I'm only going to get one knockout out
of every ten bids I send, and the time it takes to do bids is so big.' So you
just completely unblocked that constraint of time, and now you're cranking up
volume you've never been able to do with humans."

He turned it to the room: "How many of you feel limited in your capacity to
take on more work or get more clients, because the pure act of bidding, or
proposals, or trying to kick off that sales project is bloated and too big to
take on? What's cool about this society is that while his business is unique,
the problem is not. The problem is something we can all apply some of the
frameworks and thinking to, and release bottlenecks. That becomes a playbook.
And if others in the room have a similar problem to yours, Ravi, y'all can
get together and collaborate on what that looks like."

### [00:52:54] An aside — how this room gets transcribed — Austin Distel
"Also, I want to be clear: this room is being transcribed. If there are
things you would like redacted, you can always contact me and let me know.
Secondly, I already automatically redact anything that does seem sensitive —
I've built skills and agents that do redaction already. I just want you all
to be aware of it, and it will give you credit for any smart ideas that you
have. The container is safe, and it's also transcribed for your benefit — so
that afterward, all the notes and enhancements we do to this are available to
the people in this room."

He described the recording stack: "What I have found, doing this now more than
a dozen times, is one recording device is not good enough in this kind of
room. That bar is recording, and Granola is also recording. They each have
their own edge of benefit, and they're recording in different spots of the
room. When I put these together with an AI agent, it triangulates and gets
accuracy — and then we apply data on top of that, such as your member
profile, so it gets your name correct. Every tool Ravi just mentioned, from
Playwright on down, becomes a hyperlink, so at the bottom of the page you're
going to get, it will include every tool and website mentioned in this. And
one thing you'll like about Plaud is that it's the best at diarization — it
now has a frequency for Ravi's voice, so when he speaks again and I apply his
name to that voice signal, it will always catch whenever Ravi is speaking.
That's what's really rare about the other tools."

### [00:56:18] Win — Justin Day: the self-healing website system
"Just a quick one. Over the last couple of weeks I've been blowing up Fable —
thought they were going to go away, and decided to include it on everybody —
so I stayed up for like two weeks, but got a ton done. I managed to fully
configure a fully autonomous, self-growing, self-improving, self-healing
website system. That's what I'm moving towards with my local-SEO company: I
can go fully agentic with my websites and my builds with my clients, so they
get results on autopilot based on the tier they're in. I don't have to go to
WordPress or any other CMS anymore — I can completely get out of that,
leveraging the second brain I've built, collecting user data. And on top of
that, I was able to duplicate Lovable, so that if clients want to make
updates on a token basis, they can go in, just have a chat, and see it update
in real time."

For his own workflow, the interface disappears entirely: "It's fully
autonomous and self-linked — it just gives me an update on Slack for each
client on what's been done, and what needs to be done from the gap analysis
of competitors about to overtake them in Google Search Console."

Asked what powers the data layer: "I'm pulling in Google Analytics, Google
Search Console, Tag Manager, the pixel — every single part of anything I can
get from a client — and that's processing on the back end of the second
brain. Then it leverages that data to see, based on the competitors, what
pages and blogs we need to push up, using the systems and templates we
already have in place. We're not guessing anymore — we're going off of data.
That's what we've been doing as a company before, but now it's fully
autonomous. So I'm just going to scale that as hard as I can."

### [00:58:42] Win — Prashant Vanka: a personal second brain + construction dashboards
"Got two quick ones. I got a version of my own personal second brain working,
with what you guys teach. Two big things that were usually annoying: I had a
meeting with a potential LP, and they followed up with about fifteen very
legal, very specific fund-based questions — historical performance questions.
Usually I would have to fire that off to the accountant and the attorney, and
I'm getting a bill for that. After the second brain, I said: take a shot at
this. It asked a bunch of follow-up questions, flagged things that were
conflicting, I answered those — and it was very accurate." Austin: "Probably
saved you how much in accounting and legal fees?" — "Call it a few hundred
dollars an hour, times whatever they want to charge." One realization from
the exercise: "That's got to stay a personal one — it's got all my companies,
all my financials, and some sensitive information I can't allow anyone else
to prompt. But I saw the power of it."

"The second thing, on the construction side: anyone that's been in
construction — everything's over budget, behind schedule, and there are fire
drills every day. So I took the time to drop every single step of what I do,
from acquisition to final certificate of occupancy on a build, and started
building out some cool dashboards on how to manage the construction and
development company." Asked whether he runs Procore: "Right now it's all
Google Sheets. We do a lot of infill production building — high density in
super expensive neighborhoods. It's all single-family, but two or three units
in the most expensive zip codes."

He joined the Zoom room and shared the build on screen: "There are a couple of
different iterations of it. Essentially it ingests all the emails and text
messages — how people communicate right now — and surfaces what needs my
attention and what the current blockers are. Eventually, when we fund each of
these projects, aggregating all the documentation for each lender is a pain —
slightly different formatting on how they do budgets, draws, and due
diligence, so it becomes a few hours of work each time. But it's the same
stuff they're asking. I want lender packets that it automatically creates:
ingest a brand-new property, and then we just go shopping for money."

On the data-collection problem underneath it: "Fundamentally, contractors
don't want to give you project updates, go into a CRM, or use Procore
accurately. But one thing is, they need money — so I know they're entering
all the data with our lenders, and they're putting in very specific numbers
and invoices. I want to ingest information from those platforms and not ask
them to change their behavior. And then this dashboard tells me what needs my
attention and who needs a call."

The dashboard itself: "It's giving me today's chase list — all the things I
need to go address — and it prioritizes across the projects: how many days
we're behind schedule against the project timeline, how much we're over or
under budget, and what's currently on fire — go fix these." For now it's a
Claude artifact ("if you edit in it, it will overwrite when you reopen it")
and hosting is an open question he's still working through.

### [01:00:47] Win — Mike Couvillion: off Webflow, 400 YouTube videos rebuilt as pages
While Prashant was pulling up his screen share, Mike jumped in with his own
win: "Justin was sharing the website stuff he was doing. I've been wanting to
move off of Webflow for a while too — and I was a Webflow super user for ten
years, so it was a big deal for me to do that. Following the playbook Justin
shared, I had Sonnet go and scan my whole website and my whole YouTube
channel, and rebuild about 400 of my YouTube videos — transcribed all of them
with the YouTube API, built individual pages, and there's a whole search bar
so you can find all this content and context. It re-indexed all the blogs and
submitted them. This was within 24 hours of when you shared it."

The details: the new site is built to be "agentically searchable the best
way" — plain HTML via Astro — hosted on Cloudflare and published through
Vercel. "I added as much of the tracking stuff as I had on Webflow, and now
I'm just kind of like, alright, we'll see. I'll keep you all posted on what it
actually does to natural search." How long did it take? "Probably four hours
of having it run on loops and making goals towards it. Once it got a good
flow on it, it was just running on the side." He hadn't even exhausted his
weekly Fable allocation doing it: "I'm at 86% of my limit for the week, and
I've been using it on other projects too."

### [01:02:45] Sidebar — Fable vs. Sonnet vs. Opus: routing the models
Austin turned Mike's aside into a discussion: "With Fable now being the
frontier model, how are you guys differentiating when to use Fable versus
Sonnet versus Opus? Maybe our CTO in the room has the best answer, since you
do due diligence on this topic."

Mike: "One of the things I've become obsessed with is budgeting tokens. In
Claude there was always Opus Plan — it would use Opus for all the planning,
and then Sonnet for the execution, because if your specification and your
details are good enough, Sonnet can write the code as well as Opus can, which
returns costs. I've shifted that to Fable this past week. When they first
released it they said you can use it for a few credits, but they've been
upping that almost every day — as new free models come out they're getting
competition, which is good for the consumer. I'm not surprised you still have
86% left; I think they literally doubled it again yesterday."

The room traded usage tips: you can just ask Claude "what's my usage?", the
new slash-usage command breaks it down in detail, and a status bar in the
terminal can show consumption live. One member described an agent workflow
designer "that goes through a plan, specifies how the agents are going to
work together, what's the best model and the right reasoning level — and it
goes against both Codex and Claude, so it flips back and forth. It's been
pretty good at optimizing usage."

Mike described his execution layer: "I've got a farm of Codex and Claude
agents sitting there — basically a bunch of programmers waiting for something
to do." Austin: "Mike, you said you have how many AI agents working for you?"
— "I've got one server that's running about 30 Claude agents and 20 Codex
agents, and another server running a local model to do content generation. A
lot of times I'll get asked to analyze a codebase, and I can take three weeks
to step through the code and figure out how it's put together — or I can
point the thing and say, tell me what this really does in 200 words or less.
As I'm generating the workflows and the reports, I use the free models to
generate content, because that's easy and cheap and I need my tokens.
Otherwise, what I find is Codex is better at infrastructure stuff, and Claude
is better at programming and UI and front-end stuff — so it'll divvy up the
work that way. If it's going to configure a Kubernetes cluster on AWS, Codex
takes care of that; if it's going to rewrite the webpage, Claude does that."

### [01:08:53] Where this is going — dashboards in August — Austin Distel
Off the back of Prashant's demo, Austin previewed the curriculum arc: "We've
already built the Claude Crash Course — everybody here knows how to use
Claude Code now. Next is the company brain — you're experiencing that today.
Level three would be building a dashboard, because I know that as owners of
your business, insights and decisions are the most valuable thing you can do
in your company. When you get to deploy resources — people, technology —
that's what really moves the needle. Having an agentic dashboard, like what
Prashant has built, is one of those small-lever-swings-a-big-door kind of
things. So in the month of August we're going to do a training on how to
build dashboards for yourself. This has continued to surface as one of the
single most valuable things you can build with AI."

### [01:09:52] Win — Rohan Karunakaran: client dashboards LinkedIn can't provide
"Can I share a quick win on that? I run a content studio, specifically
helping business owners grow their authority, visibility, and ultimately deal
flow from LinkedIn. And LinkedIn is known for having poor analytics, poor
dashboards — and most of the tools built on top of it have non-robust,
restrictive APIs, so their analytics suck too. What I've done very recently,
as a v1, is start to build dashboarding that includes both the leading early
metrics around social engagement and the business metrics as well — no real
tool has both of those combined. It also pulls in screenshots, and almost a
narrative on how the data is flowing into the strategy — there's a 90-day
strategy. I'm able to schedule it on a bi-weekly cadence, so people just get
the download, and it provides a lot more education going into a meeting.
Meetings become a lot more collaborative. I can see a world where, once I
prove out dashboards that just work really well, this could become a software
solution at some point."

There's a sales angle too: "It's also good because I can anonymize it or
generalize it, and fire it off when I'm speaking to a prospect in a similar
vertical: 'hey, here's a recent client result, I just want to share this with
you.' There's recency, it's the same vertical, and they see actual
screenshots — 'this is the level of reporting I'm going to get if we engage.'
It's a great way to move leads along."

### [01:11:56] Company Brain training — why you need one — Austin Distel
"Our intention is that you have a simple framework, using free tools — Google
Drive — that your team can access, AI agents can access, and all the
permissions are built already into your G Suite. This isn't a new tool for
you to have to learn how to use. What's really unique about it is the way we
built the file structure so AI can read everything without burning tokens —
how to have persistent memory without filling up the entire context window.
And ultimately, your business is unique. It's a fingerprint. It's not going
to have some rigid structure. No matter if you're using PDFs, Excel sheets,
whatever — this system is still going to be able to ingest the information in
those files and build a company brain."

He set expectations for the session: about fifteen minutes of how-it-works,
Darby showing how they applied it to one of Austin's businesses, then a live
install. "I very much welcome questions. One of the values you can provide me
today is all the outstanding edge cases, questions, and concerns you have
about building a company brain in your unique business — because we're going
to be solidifying this into a company training for all members of Agentic
Society. What you're getting uniquely is participation in it, so we can solve
your unique problems too. We'll be professionally recording this training
next week."

The why: "Your AI needs context to accurately create your SOPs, run your brand
voice, all of it. Without context, it's just using the broad context it has
from ingesting the internet. Your business is unique — the way you do things
is unique — that's why you need your own brain. And it's going to grow with
you. But the intention as founder is not for it to only grow with you. A lot
of people are using things like Obsidian and other solo-player tools. I want
you thinking: how do I get my team involved and empowered by AI? And then
eventually, how do I have the AI train and improve itself? Just like you and
I are probably the kinds of people who try to become our best self every day
— what if your AI agent was building your business to become the best
business it could be, every day?"

"You don't want to be the bottleneck in your business. AZ mentioned this a few
weeks ago — he wanted to go on vacation, but his team keeps bugging him with
questions about all kinds of things in the business. He's the bottleneck of
intelligence in his company. What would it look like if your team could solve
their own problems by chatting with Claude every day, with permissioned
access to the files, tools, and information in the business — so they can
talk to the entity of the company? An entity that understands the business,
maybe sometimes even more than the founder does. Because is the founder on
every client call? Every project call? Every vendor call? Heck no."

And the equity argument: "The value of your business is not only your revenue
and your number of clients — it's also your IP. What makes your business
sellable in the future is directly related to your net worth. How valuable
are your SOPs? How valuable is your unique intelligence, and how does it grow
over time? What if all of your employees left tomorrow — would they carry a
lot of intelligence with them in their heads that was never documented, never
put somewhere, so that literally your IP is walking away? If we set up our
team with an intelligence-built system, their work silently, in the
background, improves the value of the company. And it compounds — with every
transcript, every meeting, every plan, every report, across all of your team
members, vendors, and client calls. It's getting more valuable every day with
the work you're already doing."

### [01:18:56] The three levels
"There are three levels we're going to grow into. Level 1 is what we solve
today: only for you. No one else has access. This is just you being able to
query your company brain and solve problems for yourself. Level 2 is your
team involved — where AZ can go on vacation, and his admin team can ask the
company brain questions, retrieve information, look up Loom files and SOPs
and spreadsheets and KPIs, all without having to bother the founder. Level 3
is AI agents that have full context, protocols, and an understanding of your
business — in the same way Mike has built a fleet of AI agents that goes into
a GitHub repo and understands a codebase, on behalf of venture capital and
private equity firms, to figure out in due diligence whether it's worth
buying. What if you could build your file system so that at any point in
time, I know the true value of my company, I know what every team is doing —
marketing, sales, customer success, development — and I can always ask it and
understand the status of anything? That's Level 3: AI agents that are always
improving it."

"We have to build unique features at Level 2 and Level 3 that we're not going
to build today — but the work we're doing today at Level 1 is the scaffolding
that allows Level 2 and Level 3 to happen. We're really only limited by about
one hour of install time, so let's get you a quick win today: Level 1 by the
end of the hour."

### [01:21:02] Mechanics — shared drives, scope, and the mailman
"The way this works is simple: plain MD files that review the content in each
folder. You have PDFs, you have Excel sheets, all kinds of files — the point
is not to burn all our tokens every time we want to review them. All the
security permissions are built already into Google Drive. A lot of us use My
Drive — it's kind of the trough of information everybody has — but a shared
drive is a new level of shared permissions, governance, and security
protocols, separate from your own drive. That's what we're creating today.
All your files stay in My Drive; we're going to migrate a copy of the files
you want AI to have access to."

Scope guidance: "Today's training is only for your company. We're not going to
do this for your house, your net worth, your relationships, and all the other
things you probably also have scattered in My Drive. Think about one company
— or scope it however you want. For example, we installed this for my
business The Tox. It's a spa franchise, and I own two locations — separate
entities — so I built one shared drive for my holdco that contains all the
entities in that business model. I don't have to duplicate anything; it's the
same brand and the same SOPs for both locations in Austin and Dallas, but the
permissions are separate based on who's the studio manager of each location."

"We are not going to just aim AI at all of your files today. It only has
visibility into the new company-brain shared folder you're creating, and
we'll show you ways to migrate information in — you choose what's included.
You will have duplicates; that's kind of the point. Once you have a handle on
how the company brain works and you don't want My Drive and the shared drive
holding the same information, you can retire the old copies."

Then the part he built for himself: "As I was migrating information over, I
realized I don't have the organized brain of a file-keeper admin assistant. I
don't want to think about how to rename my files for SEO inside my company
brain, or where things should be organized and categorized — I don't think
like that. So I built an AI agent framework that does it for me. I have an
inbox — 'files ready to organize' — and I just drop in anything I want in my
brain but don't know where it goes. It's like a mailman: I drop it in the
box, and the mailman puts it at the right address, renames it, categorizes
it, does all of that for me." Can you drop everything in there at once during
migration? "You sure can. And just like you'd refactor a codebase once you're
far enough into a project, this system is something you'll refactor for your
own business. You don't have to get it right the first time — that's why
we're starting in solo-player mode before team-player mode. You can just
tell Claude in natural language, 'I want permissions for this studio to be
that, and that studio to be that,' and it'll migrate the information
accordingly."

### [01:26:10] MD files, the people directory, and permissions as earned trust
"The agentic framework says MD files are the most efficient use of tokens —
beyond images, beyond PDFs, beyond Excel sheets. You want MD files that are
the fast route to directions, logs, how to think when you enter a folder. So
that looks like a README file, a company brain file, a directory, inbox
routing — these are all real, from my own company brain. My people directory
is literally my org chart: all the vendors associated with my companies, the
contacts from Gmail, our clients, our addresses — and the full directory of
all my employees, their job titles, how to contact them. Including past
employees, so it knows: I see outdated information; that person is no longer
with the company."

The folder-prompt idea: "These little MD files live in different folders
almost like a prompt for the AI agent: 'when I get into this folder, here's
our method.' How many of you do one-on-ones with your employees? Every
manager does. You have a way of doing that. You have incident reports,
payroll records — and whenever I onboard a new employee, I want a template
way to do it: a step-by-step method and a folder structure I duplicate every
single time. That's what the tiny prompt in an MD file does for you."

On permissions: "Level one is read/write permissions on folders. The next
level is you actually give an AI agent an email address and a soul file, so
it's contained as an employee — you can build a workspace account for that AI
employee and set its permissions just like anyone else's. And the way we
think about it is the way you already do with contractors versus part-time
versus full-time employees: there's earned trust over time. They earn access
over time, and you choose that as the founder."

His own tiers: "I have access to everything. My regional manager has access to
both locations, but not my financial data or legal data — that's only for me.
My location manager has access only to their studio, only their own
employees, and only their client list — my Dallas manager can't look at
information about Austin. That permission is set on the folder structure in
Google Drive, and they already do it this way — I'm just now pointing AI at
it. And then I have front-desk receptionists. I want them using AI. I'm in
the spa industry, so: 'hey, our client's pregnant — what's our protocol?'
They can just ask Claude, because it has access to the company brain and the
policies. Or 'I want to go on vacation — how do I ask for time off?' I don't
want to be asked that question. It's already a policy in the company brain.
My front desk can ask Claude, it's pulled instantly, and I don't need to
answer that ever again."

A member asked how the team actually logs in — one central Claude
subscription? "Totally. I have a Team Claude account — it comes with five
users. Dallas front desk is a user, the Dallas manager is a user, Austin
front desk, the Austin manager, and I'm the fifth user. The whole front-desk
team at a studio shares one studio email address, and in Google Drive you
share files with that address. Because they've connected their Claude account
to their Google Drive, it only has access to the folders that address has
been given permission to." Pressed on whether that's a true shared inbox or a
Google Group: "It's just a single user in G Suite — their own Gmail, their
own Drive — one individual user for multiple people at the desk."

Do you process the data into an actual database? "No, not at this time. The
processing is in the migration: it reads the file, whether it's a PDF or
whatever, and puts the most concise necessary information into the MD file.
Whenever we onboard a new employee, they get added to the directory file.
Whenever there's a one-on-one — we record all of our calls — the new
transcript is added, and the transcripts folder README gets updated: new
transcript added, here's the link. The MD file is not there to contain all
the information. Think of it like a table of contents: it reads the index and
knows, if I need that call from September, I know where to find it — but I
don't need to recall the entire transcription. You'll save so many tokens
that way. And you can adapt it: maybe the table of contents also carries the
key bullet points of what happened in each call." One member added that he
runs a similar offline index over Postgres for his infrastructure: "On any
project I probably need to know which IP address and which server — but I
don't need it all. I just say, give me the database server and the router,
and it returns exactly that."

The executive prompt: "I don't have an executive team in the sense some of you
do, so I create my own executive prompt to create department folders. For HR
I literally said: act as the leader of HR; organize your workspace — files,
contracts, reports, SOPs, plans — into this folder; list the tools you manage
(Gusto, ADP, etc.), the stakeholders on your projects, and your vendors. The
way I view this: if you onboard an AI agent the way you onboard an employee,
the first day in office you say — hey everybody, here's our new employee.
Meet the CMO, the CEO, the CFO, the CTO. Here's where you go to talk to them
about these tools and these vendors. 'Meet the CMO: their SEO agency is
this, they manage Google Analytics, Facebook ads…' That's the conversation
of the prompt sitting in every folder. When an AI agent reviews all of this,
we're basically saying: go meet everybody on the team; here's where you get
what you need to talk to them about their stuff."

### [01:38:55] Q&A — is the company brain "the harness"?
A member asked whether the company brain should be thought of as the harness —
the thing that holds the reasoning model, the skills, the prompting. The
answer from the room: not quite. The harness is how you actually drive the
model — Claude Code in your terminal is a harness; the command line, the way
it executes — while the company brain is the context and orchestration layer
the harness points at.

### [01:40:15] Live demo — pointing Claude Code at the brain
Austin showed it physically: "We all have Finder. Google Drive as an app on
your computer gets you access to all of your drives — My Drive and shared
drives. Here's my Tox company brain — this is what I installed with Darby. In
Claude Code, I just click on the company brain folder in Google Drive and
open it. Now I can switch between my companies just like this, and I'm
talking to that company brain." He asked it which employee runs the Dallas
location — and joked, watching the spinner, that he was "using Fable 5 for
the most expensive question you can ask" — and it came back with the right
name and told him exactly where it found the answer.

"Where I see this going as well: I have a bookkeeper who's a third-party
vendor, and every month she reconciles the books and sends the profit-and-
loss statement. I'd be able to query this at any time and retrieve anything
from that vendor, because it has access to Gmail and automatically pulls the
reports she's providing. Same for our Facebook ads agency or any other
third-party vendor."

### [01:43:20] Inside the brain — the router, the soul, and daily notes
Walking the folder structure: "This is my Tox company brain. It has people,
studios, an MD file that's the basic understanding of the business." What
does the CLAUDE.md do? Darby: "It's the main router — the tour guide. Claude
looks at your company brain and soul, and then knows where everything is. The
SOUL.md is the brain's personality — how you want it interacting with you,
and the really specific constraints you want it filtering things through."

Other pieces: an onboarding/kickoff prompt ("copy and paste this into Claude
Code and it walks you through the questions to gather what you may or may not
want in there"), routing and access files, and daily notes — "basically a
cron job that reviews the day like a journal. If you're trying to be your
best self, you write: what did I learn today, what can I do better? That's
what daily notes is for the business — every day it takes record of what
improved."

And the operational corpus: "HR policies, playbooks, marketing, all of our
communication, every contract, staffing, training on how to use our
scheduling tool — how do we add gratuity, how do we run sales reports.
Everything my business runs on exists in a very simple MD file structure now.
Our pricing packages — the whole reference of every service we have, what it
costs, who it's for, our ICP — lives in here, so me, my team, accounting,
finance, anybody can always reference what our official pricing is." There's
an ingestion folder for dropping PDFs, though Austin noted for simple text
policies he just has it write the MD file directly: "It's very simple text —
it's not an enriched photo or video. Maintain the record online, and write an
MD file here in the drive."

### [01:49:04] The Agentic OS MCP — install and catalog
A member asked if the MCP can be thought of as the operating system of
Agentic Society. Darby: "Yeah — and it's continually updated. We're updating
the server, writing new skills, new employees. Any time we push something
through — or a member says, 'this is a badass swarm system I want to share' —
we can process it and say: hey, Justin just released this protocol that'll
help you with building agentic websites. Now just access it inside the MCP by
talking to it." Austin: "Especially versus 'they shared that in Slack two
weeks ago — where was that thing?'"

The install steps the room followed along with: go to os.agenticsociety.com,
click Member Sign In (new members' accounts were provisioned that morning;
password-reset links were throttled for a few people, which the team worked
through during and after the session). Copy the MCP key — "it's just a long
API token" — then in Claude: Customize → Connectors → Add custom connector →
name it Agentic OS and paste the URL. "I've noticed sometimes Claude will
kick me out of MCP connectors and I just need to reconnect it — if that
happens, just reconnect. You'll see the initial tools; they're gateways into
working with the catalog. We're not going to go through the entire catalog —
you can interact with it conversationally." Darby's example of the pattern:
"Once you install the MCP once, it'll literally be: 'hey guys, we just
uploaded the coffee-Godzilla protocol.' What is that? Just ask Harper inside
your MCP, and it pulls it in, and you apply it inside the tool you're already
using."

### [01:54:39] The install — shared drive, kit, starter prompt
Then the build, live on Austin's machine, starting a brand-new brain for
Agentic Society itself: create a new **shared drive** in Google Drive (one
attendee following along couldn't see "Shared drives" in his sidebar at all —
only "Shared with me" — which the room diagnosed as a personal free Gmail
account rather than a paid Workspace account; the fix was simply to use his
company email address instead). Back on the OS portal: click **Download Your
Company Brain Kit**, and **Copy Starter Prompt**. Upload the kit zip into the
new drive folder, re-point Claude Code at that folder ("notice you're still
on your old company brain — make sure you re-point at the new one"), and
paste the starter prompt.

Darby on seeding it: "What context is in there right now? Pretty much
nothing — it's an empty folder. So this is where you'd seed it with at least
basic context about your business, or just a domain, even." Austin typed
simply `agenticsociety.com`. "That's the base-level starting point: here's my
website, let's start building out my company brain — and then it finds that
information and starts asking you questions about what you want incorporated.
If you're consolidating from scratch, a quick way is to download whatever
files you know you want it to have — treat the initial folder as a brain-dump
folder — and have Claude help you organize it." Inside the kit zip: the
canonical file structure — daily notes, people, archive, private, and the
company brain and soul documents.

The starter prompt then asks **Harvest or Interview**: "The interview is for
starting from scratch — it asks you questions and gets to know who you are.
Harvest is where it looks at the other sources Claude already has access to.
This is an important note: because you've already set up your Google Drive
connector and other MCP tools, it can reference that context — Gmail, Drive,
Notion, ClickUp, whatever's connected. If you didn't want it to know any of
that, you'd toggle those connectors off." Austin clicked Harvest — "within
seconds it knows all kinds of things from my Gmail, Google Drive, Notion,
ClickUp, and starts putting it into this Google Drive. And to reiterate:
nobody has access to that shared drive yet. It's just you. You'll give
permission access to folders within it later." (Midway through, Claude
reviewed what it was finding and announced, "Excellent — this is a gold
mine." "That's nice of them," Austin deadpanned.)

The harvest also flagged that Austin already had an existing brain in his
drive and asked what to do about it. "For now, we want this new brain only to
be for me, created from scratch. I don't want to mess with my team members'
work — I don't want them to wake up tomorrow and go, what the heck did you do
to all of our files? And the truth is, I don't know that I want all of my
employees' work in my company brain yet. So I'm going to create a new one,
and later I can go into the shared drive I already have with my team and
cherry-pick what I do and don't want to migrate."

Two practical notes from the walkthrough: if you don't want to approve every
step, you can Allow Always / bypass permissions — "the reason I did this when
I was setting up my company brain is I let it ingest content overnight. I
already knew what it was going to do. When I woke up, my company brain was
built. If you want recursive loops where it keeps improving its own content
without losing context, that's where you'd use it." And you can re-run the
starter prompt any time: "Claude will say, 'I'm analyzing your folder — you
already have this in place; maybe we want to set more things up.' I've
re-optimized and re-run mine a number of times over, and it gets better and
better." Darby underlined the ownership point: "This is owned by you. That's
why we're not teaching Obsidian — then you have another subscription and
another thing to manage. You're investing in organizing your docs, and the
value is that AI gets to use it no matter if you're using Claude, Codex, or
any model that comes out. You're not renting it — you own it."

### [02:09:05] Justin's alternate install path — gap analysis first
Justin, who already runs a deep second-brain system, described how he adopted
the kit without breaking what he had: "I leveraged what they have on the
Agentic OS — I scraped all of the different skills based off of roles, then
had it analyze my second brain, my system, to create a document I could
review internally *before* I pushed and wired everything. I could see where
the gaps were against everything I'd already built, do a comparison analysis,
build the missing roles, assign all of them, and consolidate the skills so
I'm not recreating things that would conflict. Now, based off those roles, I
have the true roles for my system, and I'm wiring it to my company brain.
Take the steps — don't do a full one-shot 'just go wire all of this.' Look at
everything on paper first. And I went an alternate direction: instead of
leveraging the MCP, I downloaded the skills to live in my company brain, so I
can run it all locally. It's working really well."

### [02:11:19] Harvesting sources that have no API — the Chrome extension trick
Austin walked through the harder harvesting problem: "Here's what I'm
thinking if I'm you: where does my intelligence live? As a franchise owner,
when I bought the rights to build these concepts, I got a huge trough of SOPs
— an LMS system with all of them. And it's a lot. Every protocol. Every
expectation of what floors to use, what desks to buy. Every policy. Every
wholesale account, every description of every SKU that's in Shopify. All of
that exists there — and there's no MCP connector; it's not some drive. So how
do I harvest that information?"

"I use the Claude Chrome extension, and I just say: create an MD file of
every single one of these policies, and put them in my Downloads folder. And
by the time I wake up in the morning, I want you to have put them into my
company drive — you can put them in the mailman inbox, and the cron job
organizes all of them. I just Wispr-Flow'd that instruction, and Sonnet goes
and does it overnight. It takes some time because it's using the browser, but
it's simple — and it's 100% going to work." Darby's alternative for the same
job: "You can provide the username and password and have it use Playwright to
go in and actually download those things locally, too."

The proof: an SOP library Austin never wrote. "I didn't write this at all.
Claude Browser wrote this, based on the content it scraped, about how to send
new employees to New York to get trained — what hotel to book, what flight
pattern. So now when I onboard a new manager, I say: hey Claude, we just got
a new employee, send her to New York — and I don't have to tell anyone the
details. It's already in the company brain."

Second technique — the inbox you already have: "Another place your data
lives is Gmail. In Claude I said: use your Gmail connector, reference the
keyword for my business, categorize all conversations, and create a directory
of all the people I've ever had conversations with — project updates, vendor
updates, employee connections, client connections. And it starts building
that out inside the company brain."

(A quick tool exchange followed when a member asked for a scraper to pull
their own YouTube videos and Instagram content into their brain for
voice-training: the room's answers were Firecrawl — "it's pretty good" — and
Exa for web-search-type retrieval.)

### [02:20:05] Beyond company data — your voice — Justin Day
Austin asked Justin how this differed from company brains he'd built before.
"These are all the right steps — it's continuing to personalize it. Gathering
all your company data is one dataset. The next step is a separate dataset
that is *your voice* and who you are: ingesting all of your Gmail — Wispr
Flow has a desktop app, so ingest that — your iMessages, everything into one
spot. Then you match your company data and your voice, and process both, so
any output has both kinds of leverage. Instead of an employee having to come
ask you questions, you pair those two things up to a Slack channel — they
just talk to you in Slack, like it's you, with the company information and
your voice. Now you've got distribution instead of people coming to you."

"And the next step beyond data aggregation is understanding how you should
*process* the data — because if you just have all the data in one spot,
you're maxing out tokens like crazy. Parse it down into actionable steps.
Each role can be its own data processor: feed the data into a role, and then
to an output."

Austin: "Reminds me of Silicon Valley, the show — Gilfoyle created his own
Slack agent of himself, and he's not typing, but he's responding to everybody
on the team, and everyone's like, what the heck?" ("Ahead of its time," the
room agreed.)

### [02:22:25] Mike's home-inventory Slack bot
Mike offered a domestic version of the same pattern: "My wife doesn't think
in terms of looking through rows and rows of stuff. She started out with just
a handful of items, so she put everything in one spreadsheet. Fast forward
two years — thousands of items, and it's unmanageable. She asked me, 'how
many of tile 7 do I have?' and then, 'couldn't you write something that would
do this?' In about four hours I had a basic inventory management system. Then
it was 'can you make it do this? and this?' — so I wrote a Slack bot that
talks to her and does it all for her. It'll even tell her, 'give me five
minutes on that — I'll come back and send you a message when I'm done.' She
has no idea" there's an agent fleet behind the Slack handle. Asked where it
runs: "I use Cloudflare for most of the hosting I do — they've got Workers
and everything, and it's the best solution for scalability I've seen."

### [02:25:17] Folder structure and self-improving routines — Darby, Mike, Justin
With Austin's live harvest complete ("inside my company brain it now has
notes, it understands people, it has skills, it has projects"), Darby shared
how he organizes his own: "Daily notes are end-of-day wrap-up stuff — a
constant ledger of what happened during the day, key decisions, anything I
want to be aware of. Clients get folders per project or venture. I keep a
folder for core skills, plugins, and workflows I create — I have Claude save
them there so I know where they're at. Same with SOPs. Keep folders specific
to what the thing is for, and label the private stuff really clearly. You
probably don't need as many folders as you think — think of them as divisions
or compartments inside the brain."

Austin asked the room: what key features would you add to make it
self-improving? Darby's daily habit: "At the end of every day, if I'm
wrapping up a thread: create a full recap, review what we covered, document
your learnings, and create a handoff document for me to jump in tomorrow.
Especially for active projects — it saves the milestones into the file
system, so if I move to another machine, or point Codex at the company brain
folder instead, it picks up exactly where we were." He also routes coding
work through the brain: "One of the locations inside my brain routes to
GitHub — 'I'm working on Project X today, go into that repo and reference
that client project for everything we're doing.'"

Mike runs a weekly self-audit: "I have an auto task that runs every week —
it's just a cron job, something that runs on a fixed interval. It goes back
through all my Claude history and tells me what I did multiple times, and
whether we could just make another cron job to automate it. It's already
pointed out two things to me this week — like, I hosted three websites on
Cloudflare and wrote people instructions for pointing their CNAMEs three
separate times, and it said: just make a skill and do that."

And a nightly one — the "dreaming sequence" he'd shared in the Slack channel:
"At night it reviews everything you did that day, creates lessons and
instincts, so that it never makes the same mistakes again, and then it
continues to progress every single day based off the lessons. You're not just
going one way — this is a loop of learning. I do a session log with
everything, and when the agents dream at night, they compact everything into
a lesson, they learn instincts, and that self-loops and grows."

He stacks a token-efficiency pass on top: "There's a skill — I'll tell it,
look at my Claude setup, look at my brain, and just iterate to use fewer
tokens. You pick one variable you want it to be better at and say: make this
cheaper in tokens; keep all your capabilities, but make this simpler. It's
literally changing one line at a time — it just can do it thousands of times
versus you typing and trying. It used to cost me about 50,000 tokens just to
float on top of context when I typed `claude`. It trimmed that down to about
7,500. It took four hours to run, and I do it about every two or three
weeks. It's simple things — you might have a 20-word line in your MD files
where you can get the same effect with eight words."

### [02:31:41] Q&A — going LLM-agnostic — Rohan Karunakaran
Rohan brought an architecture question: "I run an online agency —
operationally very light. The big thing is we're writing content, and I'm
productizing the drafting to get it 70–80% of the way before a writer goes
in. My constraint is that so much of the knowledge and context lives in
individual projects in Claude chat. And other LLMs are better at, say,
editing or drafting — it changes with every update. Right now I'm pretty
locked into Claude. I want to be more agnostic on the LLM layer. What's the
best way to set that up?"

Darby: "That's one of the limits of Claude chat — it's bucketed, and all your
context is inside it. If you use Claude Code as the main interface, looking
at a folder like the company brain — with client A, client B, client C
folders for their specific context — then you can point *Codex* at the very
same folder. Any harness can work the same brain." For running several models
against the same work, the room pointed at terminal-first tools (Cursor,
Warp): "when you're at the command line, it has access to whatever you have
access to."

Mike described his version: "I built an orchestrator that runs under Claude,
but it knows 'this is something Codex does better than me — I'll launch a
Codex agent in the background.' They can both talk to each other's config
files. I have a CLAUDE.md that's fairly generic — the rules of engagement for
every project — and another file, project.md, that's specific to the
project. When I fire up Codex I just say: go read Claude's MD file, then go
read the project file — and it sucks all that in. I sometimes forget which
one I'm in until I type a slash command it doesn't recognize."

Rohan described his current manual loop — draft in Claude, edit in Codex,
sometimes Gemini, then triangulate with his own judgment on what will perform
— and asked how to abstract it. The answer from the room: "Go to the
terminal, pick whichever one you want, and ask it: this is what I'm doing
manually — how would I automate this? They can do it step by step. I use that
hack all the time."

### [02:40:32] Sidebar — Claude Design and token efficiency
"Does anyone use Claude Design?" — several hands. One member had redone his
whole website with it but was burning through his session limits fast, even
on the max plan; the room's diagnosis was the sub-agents it spawns (you can
have it ask before running sub-agents via your CLAUDE.md) and using a
heavier model than the job needs — "I wouldn't use Fable for my website."
Mike shared a concrete efficiency win from summarizing PDF files: "It was
expensive because it was parsing all the PDF overhead, and I don't care
about anything but the text. So I built a skill that strips all the text out
first for a couple hundred tokens, then does the summarizing on that." The
generalizable advice: "Go into Claude and say, build me a skill so I get the
most efficient output for carousels — create specific things for your output
so it's not running so much it doesn't need to."

### [02:42:54] Wrapping the training — Austin Distel
"We have about 45 minutes left, and I'm glad we're getting everybody access.
I don't want this to always be a teaching, presenting kind of thing — in the
beginning, it's about connection. Y'all walked away today with an hour and a
half of solid training where you got leveled up: you know more about company
brains than most of the world at this point, in a way that's usable — not
just knowing it, but actually using it. Now I want to help you move forward
on an action step over the next two weeks, between this Friday and the next
boardroom. What would it look like for you to make a lot of progress on some
goal, process, or agent you're thinking about? You're looking around the room
— these are some sharp people. Steel sharpens steel. What would it look like
if help showed up — if somebody already had the answer to your problem, and
they happened to be across the table from you? What I ask is that you don't
just go for the problem that's top of mind because it's on your to-do list —
pick one that's actually going to move the needle in your company."

Anyone still waiting on portal access was pointed to Slack, and the recording:
"You guys have access to the recording, so you can go back through and watch
this — and keep asking questions in Skool or Slack, whichever feels more
comfortable."

### [02:45:57] Question from the room — agents for patent filings?
A member asked whether anyone had used agents for patent filings — a utility
patent for a set of technology, where the challenge is finding the claims
that are unique, aren't already prior art, and are developed in a way that
will pass USPTO acceptance. Nobody in the room had done it yet ("if I needed
to do it, I would for sure try it"). On the discovery side the room noted
Google Patents is genuinely good, and the USPTO's own search interfaces
exist but lag — "I've got all the compute in the world now, so I can work
through that. From a discovery perspective, yes. The harder part is whether
the claims are developed in a way that gets through the PTO process."

### [02:50:27] Transcript automation — the watched-folder pattern
A question about automatically routing transcripts into the second brain
kicked off the closing topic. The pattern Darby described: "Every time a new
transcript is ready, import it into a folder, and have an agent or a skill
that runs as a new transcript hits that folder — break it down; there's the
raw transcript folder and then the organized one. When we install the chief
of staff, you get that automatically. I added a skill that Roman built, I
believe, for the Zoom transcript breakdown, so you can ask it to help you
configure those. If you're recording on Google Meet, the recordings can
already go to a Drive folder automatically — same thing: watch the folder,
transcribe, break it down, and send it on. Inside that skill is where your
chief of staff watches for signals — action items, follow-up notes, urgent
items, top priorities that your team needs flagged — and it can flag people
in Slack or set up an assignment."

Austin framed the round-table: "This is relevant to everybody — how do we get
the most out of our transcripts? There are a lot of them going on all the
time: client calls, vendor calls, team calls, one-on-ones. I'm curious if you
guys have a really solid way you optimize your transcripts."

### [02:53:16] SalesBot — Ravi Parikh
"There are two things I do. I still do a lot of our enterprise sales — the
bigger contracts. And in my command line I built an agent called SalesBot.
Every sales call I do is recorded on Google Meet — I just use Google's
built-in transcription. After every call, I hit the button." He pulled it up
live on the shared screen (running in Warp — and, in true demo fashion, it
threw an error the moment the room was watching: "of course it's going to not
work when I need to demo it to you guys. I use it every day"). "The idea is:
it analyzes the transcript, classifies the call type, identifies feature
gaps and creates tickets inside our project-management software for them,
takes the call transcript and drafts an email into our CRM system, and
creates the proposal. I basically review it and click send. Creating a
proposal took me like half an hour, and now it takes me like five seconds —
and every email is fully custom, drafted off the call."

On the CRM underneath: "We're using Zoho right now, but this quarter we're
going to replace Zoho with our own custom CRM, because we're just fighting
with it. We keep getting rate-limited when we're making API calls. It's not
that expensive, but it's a pain — if we build our own, it'd be way faster and
easier." Austin marveled at the moment: "What a time to be alive. You used to
think a CRM is the heaviest moat you could ever have in software —
enterprise-grade, Salesforce-scale — and now you're talking about
vibe-coding it." Ravi's caveat: "We do have engineers making sure our vibe
coding doesn't take down our systems — we vibe-code it, and then it goes
through a QA process. That's our biggest bottleneck right now: QA."

### [02:58:06] Another post-call stack — recap, proposal, scorecard
Another member described his Fathom-driven version of the same play: after a
call, a script generates an HTML recap that goes out to the prospect with the
proposal inside it — "if it's a good call, it'll go ahead and automatically
create the payment in Stripe so they can just do it right there" — and on the
internal side it produces a scorecard of how the call went and what could
have gone better. His routing: "After the call ends, it determines whether it
was a sales call, internal, or otherwise, so it can route properly, and based
off the routing it does whatever the next steps are. Fathom has a very solid
API you can integrate directly — all my Fathom calls ingest automatically
into the company brain, and that fires off the next steps."

Austin dug into the filing mechanics: does it organize by client? "Right —
within each client's folder goes the sales call transcript with its date,
the onboarding call, any check-in calls. Clients mostly join with the same
emails, and contextually it knows what the call is, so we anchor to that —
with a fuzzy match in case they use something else." He explained fuzzy
matching for the room: "It's when somebody types something a bit off but you
still want the match — you set a parameter, 15 or 20 percent. The bigger it
is, the more open — and the more it can be wrong, so you tighten it down. In
case there's a misspelling, or a different name, or someone joins from their
phone on a different email, it still contextually figures out who it is." No
special tool required: "You set that up in Claude Code — here's my Fathom
API, I want a fuzzy match on attendees, and place it into my Google Drive
folder. Simple." (Why Fathom over Meet's built-ins? "I hate Google Meet," he
said flatly — an assessment several members co-signed, along with an even
firmer consensus that a Teams invite is where the deal dies.)

The CRM handoff, too: "It can take the data from the call and move them
along the pipeline — we use GoHighLevel as well, so it moves the card to
whatever stage. It's not always accurate, so it'll send a message on Slack
asking 'is this accurate?' for approval, and then move it. And we compact
the transcript — it doesn't live in GoHighLevel; that lives specifically in
the company brain — and add the top three notes to the opportunity, plus the
link if you ever need the full transcript. You end up with a live pipeline
where you don't actually have to do anything."

### [02:58:37] The gold-hardware catch — Mike Couvillion
Mike shared how the same transcript discipline pays off on the interior
design side, where the stakes are long memories: "For us, the sales process
is long — seven or eight months of discussion — so you have several
transcripts, and it's easy to forget that early conversation where they say
they really love or really hate something. So when we finally put the
proposal together at the end, we have the AI pull all the transcripts and
compare them to our proposal, looking for gaps or anything we did wrong.
Recently — even this week — we put together a proposal for a client, ran it,
and it flagged: the client mentioned in the second phone call, six months
ago, that they really hate gold — and the chairs you specified have gold
feet. We caught it before it went out. A five-minute change — make them
silver instead of gold — and we didn't blow the relationship before we even
started."

### [03:05:30] Build vs. buy — GoHighLevel, custom CRMs, and knowing what you actually use
Austin described where a visual automation layer still earns its keep: "In
GoHighLevel, I use the automations for things I want to *always* happen, and
I can see logs and see contacts going through. Simple things — when a Stripe
payment is made, move the contact over and start the onboarding pipeline. I
feel like it would be hard for me to manage that purely in the terminal, set
up an MCP for all of it, and review it when it breaks. There are times I
want to see a visual workflow."

Mike put the build-vs-buy question directly: "I have a related question
about GoHighLevel. I don't love it. Is there something I'm missing about why
it's so great? My client is a construction company that's not really doing
anything at a big scale — we're not using it the way it's probably meant,
and I would rather put that monthly cost toward something else. Our tech
stack for this client is so bloated — why are we paying for all these
disconnected things?" Cristian's take: if it's really being used as a plain
CRM, "you can build a CRM in Airtable or something." Someone in the room
suggested Brevo as a lighter alternative — "it does everything you need for
a small business" — with the rule of thumb that it fits companies under
roughly a couple million in revenue; Mike noted the client's business was in
a much higher band, and concluded, "alright, I'll keep it."

Austin's answer for what GHL is actually for at his companies: "I use it to
send emails at scale, track pipeline, track onboarding, and be the source of
truth for contact information — against the alternatives, it's the cheapest
but most robust option out there. And if bloat is the issue: there's an
agency view where you can turn features off per user so the client doesn't
even see them. For my spa, the front desk doesn't need automations or
reports — I just turned that off for their user and it disappears." He
showed his real setup on screen: sub-accounts per business under one agency
view; a Mother's Day promotion going to 8,000 contacts with the front-desk
team "all day just closing deals in the inbox"; a birthday automation
("seven days before a contact's birthday, send the free-facial offer — they
come in, we get the booking"); the Agentic Society member-onboarding
pipeline ("as soon as your profile got submitted, here you are — and I can
see how far you are from our activation points at 7, 14, 30, and 60 days");
and the Austin spa's lead-calling accountability board: "I require the team
to call 25 leads a day. They see the new leads, they drag and drop who did
or didn't book, and I get visibility — 'there are only three leads in here
today, all from Meta ads, and it's noon' — plus the value of the
opportunities and my conversion rates." Asked if he drives it from Claude
Code via the API: "I don't. I probably could — or should."

Ravi supplied the cautionary math for the build side: "I've done it once. It
took three months — we did it at my accounting firm, replacing the
third-party tax platform we'd been on. Three months later, we're still
finding things we didn't think about. But on the flip side, our core flow —
proposal, to collecting documents from the clients (which is most of the
work), to analyzing the documents — is a million times better than when we
were using third-party tools. With tax, everybody wants it done as cheap as
possible, so you're margin-constrained. Having the most efficient
customer-to-done tax filing is the most important thing, so that workflow
needed to be super streamlined. That's why we're doing it in our other
company too." The room's synthesis: understand the complexity you're taking
on ("those things are monsters for builders — it's way more than you think"),
weigh your priorities and actual time, and chip away rather than rebuild
what already works.

Mike added a tool watch: "Have you heard of day.ai? The Chief Product Officer
of HubSpot left to start the agentic version of a CRM. It's pretty new, but
it's purely a chat interface — everything's based on the chat."

### [03:16:56] Transcripts as a design engine — Rohan Karunakaran
Back on the transcript theme, Rohan described using them for the design
layer: "I'll have an onboarding call with a client, a few hours of content
interviews, and they'll upload a bunch of their client case studies — not
formatted, a lot of inconsistency, but they contain the actual results. Then
I ask Claude: based on that conversation, who's the client he's talking
about? Help me design a highly compelling case-study card — a one-page
visual to accompany the LinkedIn post — with the right hierarchy and the
right numbers presented. Generally designers can't come up with that layout,
because it takes knowing the context of the business and what's actually
going to make sense to a chief revenue officer. The design, layout,
hierarchy, presentation, color palette, branding guidelines — all of that is
now handled by Claude, with the context from the project files and the call
transcripts to triangulate everything." Another member added the simplest
transcript win of all: "It's great for overall discovery — go back through
the archive and find trends, or things I missed from previous calls."

### [03:18:41] The Zapier question — where the glue layer still wins
"I just met with the founder of Zapier on Wednesday," Austin said, "and he's
doing constant back-to-back user-survey calls — the thing they taught in Y
Combinator: your job as the founder is to read the marketplace; they aim for
something like thirty conversations a month per founder. He saw what we're
doing and said, you might be a good filter for how the new form of people —
AI, MCPs, APIs — are using Zapier. Where are we falling short? So: when do
you actually use Zapier now, versus two years ago?"

The room split. One member: "I don't use them at all — it must have been two
years." Another: "I love them, I use them all the time." Darby's synthesis:
"The one percent of people who are going to figure out a workaround are in
this room. Ninety-eight percent of people are not going to jimmy-rig it —
and even looking at a make.com interface will turn a lot of people off;
Zapier is a lot more user-friendly." The practical pattern several members
converged on: Zapier's MCP as Claude's fallback — "all the gaps Claude might
have, if you have Zapier connected, it just falls back to it and fills a lot
of those gaps naturally." Concrete example from the room: "Otter doesn't
connect its API to anything for exporting transcripts, so I use Zapier —
anytime there's a new one, it takes it and puts it in my Google Drive."
Another member — who'd been on Zapier's product call giving feedback before
the MCP launched — added the principle: "Not everything needs to be an
agent. It can be a workflow — a simple one-step, two-step thing. It doesn't
have to be an over-complicated orchestrated thing. That's where Zapier wins:
one step, two steps, done."

Austin's own reasons: "What I like is the visual. I see this app and that
app, the UI is clear, I'm mapping these fields accurately, and there's a log
of 'we ran a test — did it work?' I know for sure whether it's going to
work. In the terminal, you trust a lot, but there's not that verification.
And surfacing everything available for read and write is key — if you look
at the Stripe connection in Zapier, there are a hundred things you can map.
It brought up things I wouldn't have considered, that I wouldn't even know
were available through the API. I'm almost learning by connecting — like the
difference between a customer and their subscriptions, and that a customer
can have multiple subscriptions. That's not something I would have
considered just having the API ingested."

The discussion closed on a prediction: Anthropic will eventually ship the
visual workflow layer itself. "When you're building with AI, you can't
visually see the workflow — I imagine Anthropic has gotten that feedback a
lot. Some kind of canvas. They've already got everybody building MCP
connections into them, so they've hit the critical mass Zapier had to build.
They even know which endpoints get the most traffic, so they can prioritize
the roadmap." One member noted Claude Desktop "was intended to be the
graphical version of the CLI," and another wished for a live, n8n-style
view: "you can ask to see your workflow in the CLI, but it just makes you a
static diagram — I wish I could see it move."

### [03:23:04] Close — Austin Distel
"Well, guys, that wraps up the AI Boardroom. Thanks for coming — it's 12:30.
Get yourself some lunch; I'm going to Cenote if anybody wants to join me.
And I'm speaking this afternoon at Collective, sharing the Jasper story —
what that growth curve looked like and some fun stories along the way. That's
from 3 to 5 today, and you guys are welcome to come."

And the scheduling note: "Next Boardroom is actually going to be *next*
Friday. I had to shift the schedule going forward — the very first Friday of
every month is the free event downstairs, and this is the first month those
overlapped on the same day. Rather than pushing the Boardroom back three
weeks, I added one next week, and then it skips to the following two weeks —
and that's the normal cycle going forward. Your Google Calendar has been
updated."

---

*Attendee-only transcript. Please keep it within the Agentic Society.*