For the entire history of the web, the user was a person. We built sites and apps to win human attention. That just ended. The AI agent is becoming the customer, and almost nobody is building for it yet.
The old web was built to persuade a human. The agent web has to be usable by a machine. Same business, two opposite design jobs.
The user was a person. Your beautiful website existed to win their attention.
The user is an AI agent acting for someone. Your company has to be machine-usable.
Map the journey and you see it: every single step has missing infrastructure, because the old internet assumed a person was doing the work.
The agent goes looking, not the human.
Docs, pricing, APIs, reviews. Not your hero video.
Policy, limits, identity. Who is this acting for?
Paying, booking, signing, subscribing.
Filing tickets, changing settings, pulling reports.
"This one worked." Agents referring agents. The weirdest part.
Ask the obvious question. What does an agent need to be a customer that a person already had? Each answer is a company waiting to be built.
Who is this agent acting for, and on whose authority?
What actions can it safely invoke without breaking things?
Where do OTPs, docs, and threads land for the agent to read?
What does it know about my preferences and my rules?
What can it spend, and who approves the spend?
What did it see, decide, change, and buy? The audit trail.
The infrastructure layer is shipping right now. Concrete examples make the whole thing click.
Email inboxes for AI agents. Gives an agent its own inbox the way Gmail does for a human. YC-backed, reportedly doing very well.
A wallet for your agent. A purchasing agent that can buy software with spend caps, approval rules, shared payment tokens, and an audit trail.
A user's agent files the ticket, includes the logs, asks for the refund, follows up, and escalates when it gets ignored.
A CFO agent compares 12 vendors, reads the SOC 2 docs, negotiates terms, and recommends the one that fits policy.
A SaaS app hands agents real tools. Search customers, create invoices, refund orders, update tickets, pull reports. The agent never scrapes the UI.
Books the dinner, changes the reservation, calls the hotel, pays the deposit, and updates the calendar.
One version persuades a human. The other one lets a machine understand you and safely act. If you only have the first, you are invisible to the buyer that matters next.
"If an agent can't understand what you do and safely do something with you, you're basically invisible to them."
Same pages you already have. Each one flips from "convince a person" to "let a machine act."
Greg's parting gift. Take any tool you love, ask "what's the agent version," and you get a company.
His transcript ends here. What follows is Olga and Athena brainstorming against it: the deal-size question, the play Olga already ran, and the questions worth chewing on next.
↓For thirty years, every website was built to win one thing: a human's attention. Persuade the person and you win. That assumption is breaking. A second customer is showing up, and it doesn't read the web the way we do. It evaluates, transacts, renews, and recommends through structure and permission instead of brand and copy.
Greg, 645,000 subscribers, spent a whole episode landing on it and called it the next $100 billion market. The striking part isn't that it's coming. It's that almost nobody is building for it yet.
This is bigger than any one product or company. Before jumping to "here's the answer," it's worth sitting in the open questions. The deal-size one started it. The rest are below.
Olga's gut: a $20K or $100K deal isn't getting closed by an agent. Correct. The trap is picturing an agent replacing the human on the big buy. That's not the play. The dividing line isn't the dollar amount. It's how bad it is if the agent gets it wrong.
Roughly under $500 a pop. A wrong call is a shrug, not a lawsuit. The agent finds, buys, renews, and reschedules on its own.
The agent does the work and stages the decision. A human taps yes. Same trust ladder you give a new employee with a credit card.
The $20K and $100K deals. The agent reads the docs, screens vendors, and hands a human a short list. The signature stays human, maybe forever.
When you researched AI-native recruiting platforms, you did the heavy work with your agent first. Reading docs, comparing, pressure-testing, checking it was real. By the time you booked the human demo for Loxo, the evaluation was basically done. The human call was the signature, not the search.
That's the whole point Greg is missing the near-term version of. The agent-assisted buyer isn't 10 years out. It's you, this year. The agent didn't pay. It decided who got the meeting. That alone is enough to change how every site needs to be built.
Deal size was the first one. Here are the next six, the ones that decide where the real opportunity sits and who gets to own it.
Today a normal business literally cannot make a /agents surface with schemas, tools, OAuth, and checkout. So either every business hires engineers, or something generates it for them. Whoever makes any business legible to agents without a developer unlocks the whole shift.
The Loxo story says the assisted buyer, a human steering an agent to research and shortlist, is here today. Autonomous agents spending money is later. So the first dollar is probably "be discoverable and trusted when someone's agent does the research," not "let agents check out."
Agent-invisibility is silent. You never see the deals where an agent screened you out and moved on. Whoever makes that loss visible, which agents came, what they asked, where they bounced, gives the market its reason to act.
Counterintuitive answer: the small, local, transactional ones. Bookings, deposits, reschedules, restocks are exactly the low-regret transactions agents will run end to end. Main Street may get reshaped by agents before the enterprise does.
When a buyer's agent shows up, most businesses have nothing on the other side to answer it. A growing asymmetry: armed buyers, unarmed businesses. Who arms the seller side, and does a business get its own agent that quotes, answers, and books?
Brand, reviews, social proof. An agent ignores all of it. It checks identity, policy, limits, receipts. So what becomes the new trust layer for machines, and who gets to set the standard everyone has to meet?