Warning: Investors Must Prepare For The Agentic Takeover

An "agent" is AI that does tasks for you — booking, comparing, buying — instead of just answering questions. Meta's new Muse app is one, and Carlson thinks whoever runs the agent people trust becomes the new front door to shopping. Once it remembers your preferences, your family's allergies and your saved logins, every future request gets easier, and that habit is hard to break. Meta could then charge for sponsored suggestions inside the agent or take a small cut of purchases.

He also thinks Meta's existing ads are safer than Google's. Google earns money when you search for something you already want, and an agent can do that searching for you. Meta earns money by showing you things while you scroll for fun, which creates the want in the first place — and people will keep scrolling Instagram even with an agent. He owns Meta and has been buying a lot of it.

Citrini's essay predicted Mastercard would suffer. Carlson thinks the opposite. Mastercard doesn't care who owns the shopping app — it earns money on the payment itself. When software agents start paying on people's behalf, new questions appear: was the agent allowed to spend this, was it within the limit, can the charge be traced or reversed, is this a real agent or a fraudster? Mastercard's tokens (stand-in card numbers) and authentication tools answer exactly those questions, so more agents mean more need for Mastercard. It is one of his biggest holdings, about $189,000, and he keeps adding.

Inside a company, an agent needs to know who is allowed to see which files, who can approve spending, and how every action is logged. Microsoft already controls that layer — company logins, email, documents, meetings, cloud and security — so agents at work will most likely run through Microsoft. That makes it a likely winner rather than a victim, even though software stocks have been weak. He holds it.

ASML makes the only machines that can print the most advanced computer chips. Agents don't remove the need for chips — they use huge amounts of computing, which means more chips and more ASML machines. No AI can recreate decades of precision engineering through a better app, so he calls ASML "the apex predator" that agents can't touch.

Visa gets the same argument as Mastercard: a card network is something merchants and shoppers must use, not a convenience an agent can skip. Its plan for agent shopping — tokens, identity checks and spending limits — makes it part of the trust layer that more automated purchases need. He owns Mastercard rather than Visa but groups them as likely winners.

Some investors fear AI will replace S&P Global because AI can write research summaries. Carlson thinks that gets it backwards: those summaries are built from S&P's data, so AI raises demand for it. Credit ratings are also required by regulators and written into loan contracts, and a bank lending billions needs a trusted, checkable source, not a cheap website. The value comes from authority, which an agent can't copy.

An agent might tell you what to watch across all streaming services, which could weaken Netflix's own recommendation screen. But Netflix owns or licenses the shows themselves — the agent can point at a show, but it can't legally stream it without Netflix. So he sees Netflix as very insulated.

An agent can compare prices, but it can't copy what makes Costco cheap: its buying power, small selection, own-brand quality and membership model. A price-checking agent may even send more regular shopping to Costco because it can see the value. Costco is a supplier, not a comparison website, so it is resilient.

An agent can book a table, but it can't cook the steak or recreate the atmosphere. Because Texas Roadhouse's restaurants are so well run, he thinks agents that make booking easier could bring it more customers, not fewer.

Google is his largest holding, and this is a mixed case. Google earns its search money when you run several searches and click ads along the way — "best hotel in Chicago," then a few more. An agent might do all of that in one step and hand you one or two answers, leaving fewer chances to show ads.

But Google has more defenses than most: Android, Chrome, YouTube, Gmail and Maps reach billions of people; it knows what people are looking for; it has its own AI (Gemini) and cloud; and it is building the standard that lets agents check out with merchants. The ad could become a paid spot among the options the agent considers. So he sees Google as much safer than the truly exposed names.

Many people order from DoorDash out of habit. An agent that compares DoorDash with Uber Eats, ordering direct from the restaurant, or pickup breaks that habit. But delivery isn't a website lookup — it needs thousands of drivers nearby, smart routing, fraud checks and refunds. Whoever has the densest network will still win the agent's order, and that's likely DoorDash. He holds it, keeps it, and thinks the recent dip reflects this risk being priced in.

An agent could compare Uber with Lyft, Waymo and Tesla every time you need a ride. But it will pick the most reliable, fastest and safest option at a fair price, and Uber's huge number of drivers usually makes it that option. The agent can talk to Uber's app, but it can't replace Uber's network.

Amazon has been the default place to start shopping. An agent that searches every website at once weakens that, and could hurt the paid product listings Amazon earns from. But Amazon also has the huge selection, Prime, fast delivery and easy returns an agent would want for a reliable purchase — plus AWS, which can run the agents themselves. Some of the business is hurt, some helped.

If an agent can read and shop in any language, you need less language skill to get things done. But most Duolingo users aren't learning to buy things abroad; they want to talk to people and connect with a culture, and they enjoy the streaks and progress. That motivation isn't something an agent replaces, so he doesn't see Duolingo as highly exposed.

Booking (with Kayak and OpenTable) exists to gather travel options and sort out the terms for you — exactly what agents are good at. It already spends about 30 cents of every revenue dollar on marketing, largely buying Google ads to catch travelers. If people start telling a personal agent "find me a quiet hotel under $350 with free cancellation," the agent, not Booking, owns the customer.

Booking doesn't have to lose bookings to be hurt: it could become a behind-the-scenes supplier that the agent picks from, paying the agent instead of Google or accepting a smaller cut. Unlike DoorDash, it has no fleet of drivers to fall back on. He sold it, and he's more worried now than when he did.

TurboTax is basically a long, friendly form that walks you through your taxes, and people pay $100–200 for it because mistakes are scary. An AI agent that already knows your finances and can read your documents could do most of that quietly all year long, making a separate tax product harder to justify. Credit Karma faces the same squeeze. He sold Intuit and is now much more concerned for its shareholders than when he sold.

LegalZoom sells standard legal paperwork — forming a company, simple contracts, wills. Those are templates, and AI can already fill them in and explain them. What stays valuable is real legal judgment, which is not what a template service sells. So he puts it among the most exposed.

Zillow's value is helping you search and compare homes. An agent can already search listings, compare neighborhoods, book tours, watch for price drops and gather documents for you — so the reason to open Zillow shrinks. He sees it as very much at risk.