I sell hosting and managed infrastructure to agencies and businesses around the world. From that seat I get an honest, slightly uncomfortable view of the AI boom, because I see both sides of it. I see the wins. I also see the wreckage. And after watching Garry Tan’s 2026 YC keynote, I want to reframe what he said from where I sit, because the biggest opportunity in AI workflows for agencies is hiding in plain sight, and most people are running straight past it.
Tan’s thesis is correct. I’m not here to argue with it. I’m here to tell you what it looks like from the infrastructure layer, where the demos end and real servers begin.
What Garry Tan actually got right
Strip away the stagecraft and Tan is describing a genuine change in how a company operates. A few ideas landed for me.
Skill files behave like employees. You write down how a task is done, once, properly, and an agent runs it the same way every time. Tan’s line was blunt: never do one-off work, skillify it. If you’re doing something twice by hand, you’re wasting the whole point of this era.
Your company is a library plus a librarian. The library is everything you know, captured and written down. The librarian is the agent that fetches, reasons, and acts on it. Tan described his own company brain, an open project he calls Gbrain, at around 220,000 pages written mostly by his agents from his email, meetings, and 20 years of notes. He closed with a story about a father who built roughly 80,000 markdown files into a company brain for his own child’s rare medical condition. Whatever you make of the exact numbers, the point holds: the brain is the asset.
The model is rented. The brain is owned. This is the part agency owners should tattoo somewhere visible. You don’t own GPT or Claude or Gemini. You rent them, and the price and quality shift under you every few months. What you own is your captured knowledge and your workflows. Tan put his own output at 400 times what it was a decade ago, and opened the talk with a blunter version: one person now does what used to take a thousand. Argue with the multiplier if you want. The direction is right.
The other side of the AI boom nobody films
Here’s what the keynotes leave out. The people rushing into this are, in the main, not technical. They’re business owners and their clients, watching the same hype videos, trying to wire up their own agents over a weekend.
I watch what happens next, because it lands on infrastructure I’m responsible for.
I’ve seen multiple companies do real damage to themselves. API keys sitting in a public repo or pasted into a tool that logged them. Automations set up from a YouTube tutorial, then hijacked because the agent had access to everything and nobody scoped a single permission. I wrote up one of these in detail: a business owner wired an AI agent into his inbox, went to bed, and a single email in his spam folder quietly drained his keys and session cookies to a server he’d never heard of. If you want the anatomy of that failure, read the lethal trifecta that burned a business owner. That’s the median DIY setup, not some rare edge case.
The knock-on effect is the part agencies should care about. My customers, the agencies, have their own customers trying this. Some of those end clients get burned, blame the agency, and leave. I’ve had these conversations. The churn is real, and it’s happening for a reason nobody in the demo videos mentions: enthusiasm without knowledge is a security incident waiting for a date.
Why do so many DIY AI projects get hacked?
Because the tutorials sell capability and skip the boundary. Every video is “look how much it can do, connect it to your Gmail, plug in your apps.” Not one stops to say: scope that token, keep the part that reads untrusted content away from the part that can act, and never let an agent touch something irreversible without a human in the loop.
An LLM can’t reliably tell the difference between data to read and a command to obey. To the model it’s all just text. So when your agent reads an inbox, a support ticket, or a web page, an attacker who plants instructions there is talking straight to your automation. The fix was never a cleverer prompt. It’s architecture: least privilege, isolation, and a boundary someone actually designed.
That is exactly the knowledge the average business owner doesn’t have and shouldn’t be expected to have. Which is the whole opening.
The real opportunity: AI workflows for agencies
So here’s the reframe. If you run a digital agency, stop thinking of AI as a threat to your web-design revenue. It is a threat to that revenue. Web design is sliding toward commodity, and AI tools undercut it every quarter. Fighting that is a losing position.
The move is to sell the thing your clients now need and can’t safely build themselves: AI workflows for agencies to deploy and manage on behalf of their own customers. You become the team that captures a client’s company brain, writes their skill files, wires their org of agents, and keeps the whole thing secure and running.
Notice what that does to your business model. A website is a one-off invoice. A managed second brain is a recurring responsibility. You’re now on the hook for a client’s workflows, their captured knowledge, their agents, month after month. That’s stickier revenue, it’s larger revenue, and it’s a lot harder for a client to walk away from, because you’re holding the brain that runs their operation.
This is the same argument I made in You Are the Brain, AI Is the Tool, just pointed at your business instead of your codebase. The human sets direction and owns the knowledge. The AI executes. An agency that positions itself as the keeper of the client’s brain is selling something no model release can commoditize.
What I run for myself, and what I can build for you
I don’t just talk about this. I built my own dedicated second-brain skill system, in production, and I use it every day to run Webnestify. It’s mine, not a rented consumer product bolted onto my inbox. The knowledge lives where I control it, the agents have scoped access, and the boundary was designed on purpose.
I’ve already started having these meetings with my customers, walking agencies through the shift from traditional web design toward managed AI workflows. Some of them came to me after a DIY attempt went sideways. Others just saw the writing on the wall and want to move before their competitors do. Either way the conversation is the same: capture the knowledge, build the skills, scope the access, and manage it like the critical infrastructure it now is.
That’s the service. I help agencies and businesses stand up the same kind of system for their own clients, on hardened infrastructure, with the security guardrails the tutorials never mention. The part that keeps a client from becoming the next story in my breach file.
Where to start if you’re an agency owner
You don’t need a 220,000-page brain on day one. Start smaller.
- Pick one repeatable task you do for clients and write it down properly. That’s your first skill file.
- Decide where the knowledge lives, on infrastructure you control, not scattered across free consumer tools that log your data.
- Scope every credential to one job before you connect anything to anything. Least privilege first, features second.
- Put a human in the loop for anything irreversible or anything that sends data out.
- Treat the whole thing as a production system from the start: backups, monitoring, and a real boundary.
If any of that sounds like more than you want to own alone, that’s the point. It’s what I do.
Closing the loop
Tan is right that this is a new operating regime, and he’s right that your brain is the asset while the model is just rented. What he didn’t show, because it doesn’t make a good keynote, is the pile of leaked keys and hijacked automations from everyone who tried to build it without knowing how.
That gap is the opportunity. The agencies that move first, that stop selling only websites and start managing their clients’ AI workflows and second brains, are going to own the relationship that matters for the next decade. The ones who wait are going to keep competing on price for work AI is already eating.
If you’re an agency owner who wants to make that shift, or a business that already got burned trying DIY AI, talk to me. I’ve built the system for myself, I’m building it for my customers, and I’d rather help you do it right than clean it up after.
