Anthony Conrad @Revenintpro

ChatGPT had this chance first. We've been through its app review. Test cases, policies that keep changing, and no updates allowed while you wait for a review that can end in a rejection. Muse's connector platform asks only for what an agent needs to verify. The intent from Meta looks promising. And Meta is best placed to build the ad layer on top, with connectors competing to get picked. Agent discovery will need its own ads.

Was talking to a friend the other day. He's built a B2C company and scaled it like crazy. We got into the usual game. What still makes money five years from now? We kept landing in the same place. If a business is only digital, it's going to be hard to defend. An agent will either rebuild it or just go around it. The ones that last have something physical, or a human doing the work. Cabs. Food delivery. Even the visa agent who sorts out your paperwork. Then I asked what he'd build if he started today. Either robots, or an ops-heavy business like a real estate marketplace. Agentic only in how customers find you. The ops underneath stays yours. Robots sounded far off until I looked at Waymo. 50,000 paid rides a week in May 2024. 500,000 by March this year. That's a cab business, and it's growing faster than most apps. The second idea is where it gets tricky. In these ops businesses the real moat is the data. Which listings are live, what people actually pay, who converts. The moment you go fully agentic, you hand that to whoever runs the agent. So a lot of them will hold back. You can already see it. Last week Amazon blocked Meta's Muse from shopping on its site. Amazon did over $68B in ads last year, and all of it depends on people scrolling its pages. An agent that just buys the thing doesn't scroll. Zillow went the other way, but carefully. It's inside ChatGPT, but OpenAI can't train on its listing data, and every tour and lead still goes through Zillow. So I don't think the question is whether these businesses go agentic. They'll have to. It's how much they're willing to show. My guess is the ones that win let the agent in the front door and keep the back office to themselves.

ChatGPT ads hit a $1B run rate in 200 days, per Graphite and Similarweb's new study of 98k ads. Mostly US data. We captured the India version: 279 ads from 104 advertisers in 13 days in September. - 66% of advertisers ran a single creative. Jotform alone runs close to 1,000 in the US. - Travel took the money. 28% of travel answers carried an ad. For corporate travel, 48%. - In 95% of ads, the advertiser wasn't named in the answer above it. Birla Opus bought every paint slot we saw. Asian Paints got the recommendation.

Gokul is right that DoorDash doesn't have to hand food and grocery to Muse or Instinct. The moat is real. Nobody else is getting hot food to my door in 10 minutes. But look at where the profit comes from. Swiggy said it plainly in a shareholder letter last October: more than the entire profit of its food delivery business comes from ads. Every time I open Swiggy or Zepto, the first things I see are sponsored. Brands pay for that screen because I'm about to buy. Now picture the order coming from an assistant. I say restock the kitchen, it calls the app, the order lands. Nobody scrolled. Nobody saw the banner. The app still delivers in 10 minutes. It just stops being where the buying decision happens, and the ad budget follows the decision into the agent. That's why Amazon blocked Muse. Blocking protects the discovery, not the delivery. So the question for vertical apps isn't whether they keep the category. It's whether they keep the screen.

At @apify's Agentic Commerce Builders Night (#SFTechWeek), the best talk was Matej Hamas on how Apify got AI agents to pay for its tools. The problem: a shoe has a fixed price. A scraper run doesn't. You only know the cost once it finishes. What they tried: 1. Skyfire. It worked, but had little traction. 2. x402 "exact" plus a refund afterwards. Hard to guess the upfront charge, and the refund sits outside the protocol. 3. x402 "upto", which caps the charge. Risky, because the agent's funds may be gone by the time you charge. 4. Their own agent gateway with prepaid tokens, over x402 and Stripe's MPP. 5. An escrow model launched that day. The agent locks up to $X and pays for what it actually uses. About $100 a week in agent payments so far. Early, but the plumbing is getting built.

Claude Founder House, day 1. The line went out the door, but I finally got in. Claude made me this card: "The Connector. Half the room is here because you texted them." Fair. I'm in SF through October building t.co/VGJTYXgBQq, which gets brands ready for AI agents to shop. If you're working on agentic commerce, text me too.

Two weeks ago Amazon blocked Meta's Muse for shopping on Amazon without permission. Yesterday Meta teamed up with Sierra and announced PAP. I don't think that's a coincidence. PAP (Personal Agent Protocol) is basically a standard way for your personal AI agent to log in to a business and do things for you. You decide if it can only read or actually transact. It's just announced, not launched yet. v0.1 spec lands end of this month. But it looks promising. My first thought was, why Sierra and not anyone else? Then it clicked. Sierra is the one selling brands their own agents. And agent auth is the biggest challenge businesses face when they want their agents to actually transact. And with Bret being Facebook's former CTO, the Meta link makes sense. The way I see it: - customers are starting to show up as agents, not people - logging in and getting permission is the missing piece - every agent-to-agent conversation pushes more business to Sierra - and it's a land grab. UCP, ACP, now PAP Also notice who's not on the list. Amazon, OpenAI, Anthropic. That Muse block is exactly why this exists.

3 days at Claude Founder House in SF. Jared Kaplan, Mike Krieger, Netlify's CEO and Lovable's growth lead on one stage. The line everyone repeated: build for the model that ships in 6 months, not today's. If your product is perfect now, you're not building far enough ahead. t.co/m0Hrehguj0

Mike Krieger (@mikeyk): build for the model 6 months out. Ship with heavy scaffolding today, snap it off when the model catches up. Like supports on a 3D print. His 12-month call: disposable software. Interfaces get thrown away. Data, relationships and connections persist. t.co/tF2M2wF8dV

Met @gregisenberg at Claude Founder House yesterday. Told him I used to read Idea Browser every week. Then I picked an idea, started building and stopped opening it. Then we talked about what a newsletter for people who already picked would look like. I want it. t.co/NX01DX8hsE

Read the skill this installs, not the landing page. It’s a marketplace with the front end removed. Shop’s catalog across millions of stores. Shop’s ranking, which the agent is told not to touch: “buy me a moisturizer” goes in as “moisturizer”, not “well-reviewed moisturizer”. Your addresses and saved card, Shop Pay, order tracking. The website is now whichever agent you already use. Returns aren’t in it. Two things the page doesn’t say. Off Shopify, the skill finds nothing for you. It issues a single-use card so the agent can pay at any store, so off Shopify, Shop is the wallet, not the shelf. And nothing in it compares a brand’s store with Amazon or Target. That one view across all three is still nobody’s product, and it’s where the demand goes next.

Robot foundation models are becoming an API with a fine-tuning loop. Almond trained a deburring policy in four days. The bottleneck is shifting from models to data collection.

At @vercel Ship SF, @rauchg on stage. Shipped today: Python GA with containers and daemons, Vercel Drives for stateful agents, vercel.store on Shopify Hydrogen and eve, BYOC on AWS, and scoped per-task agent access through Vercel Connect.

Did some research today & found a supplement brand with decent numbers Guess their payment processor 3...2...1... @whop coming for the full ecom space

i'm sorry, Jev can WHAT? it can take your website link, find people who need your product, email them and book the calls? since when?? i've been doing this by hand for a YEAR. ok. testing it: $30 free 👇 explee.com/auto-gtm/x/tw-…

The craziest demo I saw this week was an AI agent negotiating a software procurement contract with another AI agent over email. They haggled over the SLA terms for three hours and finally settled on a 15% discount. The future is weird.

Stop stressing over multi-currency pricing. Stripe just rolled out instant currency conversion in the dashboard for 13+ currencies. No more getting killed on exchange spreads when you repatriate your EU revenue. 💶💷 #Stripe #Ecommerce #SaaSMetrics #RevenueGrowth #PaymentProcessor

Vibe coding is so 2025; the new meta is vibe deploying. Between Stripe Projects and the new embeddable Checkout UIs, you can provision your entire billing infrastructure from a single terminal prompt. The developer experience is unmatched. 💻⬛ #Stripe #DevTools #SoftwareEngineering #StripeCheckout

Your AI is only as smart as the APIs it can talk to. An agent that can write a beautiful email is a toy. An agent that can read a contract, query the CRM, draft the invoice, and execute the payment via an API is a business.

Cross-border payments are finally solved. Stripe integrating Bridge and launching the Tempo stablecoin chain means we can pay international contractors in USDC instantly without the ridiculous FX fees. Crypto rails, fiat UX. 🪙💸 #Stripe #Stablecoins #GlobalTeams #StripeCrypto