The website is dead: inside the coming agent economy

“I am convinced the website is dead”

The compilation opens with a claim Muzamil is willing to plant a flag on. “I am convinced the website is dead,” he says, before pushing further: “Your ad will never be seen by an agent.” The framing is deliberate. This is not a prediction about a marginal shift in how people search. It is a prediction that the surface most of the consumer internet is built on, the rendered webpage a human scans with their eyes, stops being the surface that matters.

He grounds the claim in a generational tell. He pictures explaining to his kids the way he used to book an Airbnb: typing into Google, clicking into a website, pasting links into a spreadsheet, sharing it with a friend for feedback, then finally booking. The reaction he expects is blunt. “Oh my god, what is this? This looks medieval.” The point is not nostalgia. The current flow is already indefensible once a better one exists.

The pizza test inside Claude

The strongest piece of evidence in the episode is a small test Muzamil ran that same afternoon during a customer demo. He asked Claude to find pizza delivery to his office. What came back was not a list of search results. “It literally gave me three connector options,” he says: Uber Eats, Glovo, and one other.

Then the line that carries the whole thesis. “The amazing thing is it gave me the first answer without doing any web search. It did not go on the internet.” The assistant looked at its connector registry first. If Muzamil said no, only then would it fall back to the open web. If he said yes, an MCP connection would pop up and complete the order. The default path never renders a page, never shows an ad, never rewards SEO.

The optimization layer always shows up

Muzamil places this inside a pattern anyone who has worked in distribution will recognize. Every merchant platform grows an optimization layer on top of it. “You had Google, you had SEO. You had the App Store, you had ASO. You had AI so far, which is chatbots, which is AI citations and AEO.”

The new entrant in that lineage is the connector itself. OpenAI moved first with ChatGPT apps, which later merged into MCP, and MCP has since been adopted broadly enough to function as a standard. A business can build one integration and plug into a large number of platforms at once. For the ecosystem this is good. For incumbents whose moats were built on being the destination, it is not.

Applications become infrastructure for agents

The subtler point is what these applications become. They are no longer just the last-mile experience a human sees. They are also the infrastructure an agent reasons over when it decides what to use on the user’s behalf, as Muzamil puts it, “without letting the human know.”

That reframing changes the product question. The winners today, he argues, are “the ones that are leaning into this thing and trying to stay ahead of the curve,” the ones “opening everything up first” and going MCP-first. They make it easy for agents to extract value from the product even when the user never actually logs in or opens the app. The old instinct, force the user into your surface so you can monetize attention, becomes actively harmful when the user has delegated the decision to an assistant that will route around anyone who resists.

Agent-to-agent as the endgame

Muzamil pushes the picture one step further. “You will have your AI agent. Your AI agent knows everything about Muzamil.” It knows when you want to go on vacation, when your family is planning a trip, the full context. That agent will build the itinerary, and then, in his words, go say, “look, I want to go talk to the agent of X travel search company.” The exchange is AI to AI.

The precondition is infrastructure. That handshake can only happen, he notes, “if the underlying infrastructure is able to offer that breath.” This is why MCP-first is not a marketing stance. It is admission into the next distribution layer.

A $30,000 safari, booked end to end

To keep the argument concrete, Muzamil surfaces a live deployment. In a travel product called Layla, an AI runs an entire department. Users converse with Layla and end up doing things like booking a South African safari for their kids. The AI understands the pain point that would have sent them to a travel agent, generates a video addressed to that pain point, produces six different formats for TikTok, Instagram, Google, and YouTube, deploys them, runs A/B tests, has access to the internal tools, monitors which tests bring in new users, and loops.

“That was pretty shocking for us to see,” Muzamil says, “basically that it can handle so much context and react so fast in real time.” The scale he cites: around 5 million users, roughly 30 million messages on the platform, and about a billion dollars of trips built inside the product. The proof point that lands hardest is the most recent one. Just the previous week, a family booked a $30,000 safari end to end through the AI, with the human only needing to press confirm at the very end.

At that price point, the argument that people will always want a human in the loop for big-ticket purchases looks weaker than it did a year ago.

The counter-thesis on Google

The episode does not end in a straight line. Muzamil gives space to a genuinely contrarian read on Google, one he says runs against the surface consensus that Google is being disrupted. As the LLM engines scrape “all of these pages, or actually the database built on top of Google, or all that information that is being fed on top of it,” the LLM providers are drawing from a base Google largely owns.

His argument is about what happens after the current funding cycle. Right now, LLM companies are burning capital at massive scale. Once that settles, Muzamil thinks Google still holds meaningful leverage, because what the LLMs eventually show users is “some relevant most relevant links that are clickable and taking you to some pages,” or a solution inside the chatbot that still traces back to indexed content. The interface changes. The underlying data gravity may not shift as fast as the headlines suggest.

Vanilla with sprinkles is still vanilla

One line from the opening keeps resurfacing across the episode. Muzamil warns against companies that treat AI as decoration on an existing product. “No matter how much you try to add, you try to add some chocolate sprinkles, um rainbow sprinkles, add some strawberries on top, it is still vanilla at the end of the day.”

Read across the compilation, that image is the throughline. The interesting question is no longer whether AI will change consumer software. It is which layer of the current stack survives the transition, and which gets absorbed. The agent, in Muzamil’s framing, is not a feature bolted on top of the app. It is the new front door, and everything behind it, connectors, MCP, agent-to-agent handshakes, is being wired up right now. The businesses that treat this as a UX refresh will lose. The ones that treat it as a distribution reset are the ones the show keeps circling back to.