AI agents break web math, zombie UI, and Harvey's margin collapse

AI agents break web math, zombie UI, and Harvey's margin collapse

The Throughline Brief · September 27th. 6 min 26 sec.

Cold open

Amazon's core retail business barely breaks even. They need their 70 billion dollar search ad network just to survive. Now imagine what happens when AI agents start buying things for you. They skip the search bar. They skip the sponsored products. They go straight to checkout. The ad revenue vanishes. That's a margin collapse waiting to happen.

Intro

This is The Throughline Brief for September 27th. I went through every major AI and tech podcast this week so you don't have to. I'm West. Here's the stuff that actually matters for builders and operators. Let's get into it.

Agents Break the Web

For twenty five years the internet ran on a very simple exchange. Users get answers and publishers get impressions. That math is completely breaking down. Parag Agrawal went on 20VC with Harry Stebbings and laid out the exact sequence. An AI agent queries an index. It extracts the text. Then it returns a synthetic answer directly to you. No human in the loop. The agent never views a display ad. It never clicks an affiliate link. Web publishers lose their only real revenue stream. Agrawal put it perfectly. Agents show up, no one sees ads, you make no money. Think about that. The consumer app side is totally exposed too. Chamath Palihapitiya talked about this on the All In podcast. He pointed out that headless agents skip visual interfaces entirely. They turn walled gardens into raw utility layers. An agent booking a flight or clearing out your inbox does it all behind the scenes. It strips away the application storefront. That threatens Apple's famous 30 percent fee. It forces massive subscription platforms to compete purely on pricing efficiency. Then there's the corporate brawl over traffic. John Coogan brought this up on TBPN. He noted that Amazon will aggressively block external agents to protect its margin. Amazon retail makes almost zero net income without its ad business. Third party commerce agents threaten to wipe out impulse browsing by jumping straight to checkout. The war for this traffic will be absolutely brutal.

Margin Collapse

Let's talk about unit economics. Multi step agents frequently fall into endless reasoning loops. Those loops drain your compute budget incredibly fast. Claire Vo broke this down on How I AI. She noted that most agent setups collapse after just six or seven autonomous tool calls. An autonomous system hallucinates its state. It repeatedly queries an API without ever actually resolving the task. That racks up massive unintended inference bills. It completely breaks fixed cost subscription models. Context drift sets in. API tokens fill up with intermediate garbage. The whole system loops endlessly. That destroys unit economics for application builders. John Coogan tracked a brutal example of this margin collapse at Harvey during a discussion on TBPN. Lawyers stopped asking single questions. They started letting agents run autonomous legal research across firm wide document stores. Token consumption went absolutely vertical. Harvey's gross margin fell from about 50 percent to negative 50 percent by June as agent token use spiked twenty fold. This is a massive trap for founders building in this space. Selling fixed price software seats while paying for open ended reasoning loops creates an uncapped liability. You can't sell software the old way if the compute cost scales with the length of an AI hallucination. The numbers simply don't work. Every operator needs to audit their pricing model right now.

Danny Meyer

Danny Meyer talked about startup culture on How I Built This. He explained why calling your company a family creates massive toxic debt when you finally have to enforce standards. Here's what he said. A business isn't a family. Every time we held someone accountable people would go into deep grief because you don't fire a family member. That rings so true for early stage teams. You have to drop the familial rhetoric immediately. Focus entirely on explicit behavioral standards. When you blur the lines between a team and a family you make it impossible to give honest feedback. Build a high performing team instead.

Claire Vo

Claire Vo brought up a hilarious truth about model preferences on How I AI. She explained why developers actually walk away from frontier models after using them for a while. She said it plainly. I stopped using Claude not because of its intelligence or its models. I stopped using Claude cuz it was annoying. This is exactly what builders miss when they obsess over benchmark scores. The real reason engineers abandon a tool is usually preachy verbosity. If your agent lectures the user or pads every answer with five paragraphs of filler they'll leave. Real human utility beats raw intelligence every single time.

Molly Graham

Molly Graham sat down on Lenny's Podcast to discuss delegating work to autonomous agents. She pointed out that outsourcing production to a machine demands an incredibly sharp understanding of quality. She framed it like this. If you don't know the definition of good, how can you give it to your summer intern. If you don't understand what quality is, you can't outsource that yet. Here's why that matters to you. Delegating work to an AI agent doesn't absolve you from owning the final outcome. The software can generate the draft. You still have to provide the taste. Taste is your ultimate moat.

The bottom line

The big takeaway this week is incredibly clear. Software generation is practically free now. We have unlimited words and unlimited code at our fingertips. But human taste and solid unit economics are rarer than ever. You have to nail your pricing models before inference costs eat your margins. And you have to protect your craft. That's your Throughline. See you next Sunday.