Stanford MS&E435: Who Chooses the Technology Stack—the Developer or the Coding Agent?

“I don't even know what Vercel is. My agent brought me here.” According to Guillermo Rauch, this is what a new customer who encountered an error wrote to him in an X direct message. One sentence contains the new economics of software distribution: the person did not compare clouds and may not even have known the platform's name. Their favorite coding agent chose the deployment platform :)

This is another review of a lecture from Stanford MS&E435, where we began with a map of the economics of the AI supercycle and have now reached applications and the question of who captures value when software becomes more abundant. Guillermo Rauch, Vercel's founder and CEO, offers this answer: code is becoming cheaper, but the result still has to be deployed, run, and maintained. Vercel's economic bet is to occupy the layer between intent and a working product: give the agent an API for deployment, preview URLs, an isolated environment, a gateway to models, and long-lived processes—and eventually a cloud that diagnoses operational problems itself. It is both a substantive lecture and a carefully constructed vendor pitch.

What happens before deployment is more interesting. According to Rauch, an agent arrives with an existing “world model”: models have already learned about Next.js, React, and Vercel's open-source tools from material on the internet. Such an agent does not conduct a neutral analysis. I would frame the economic consequence this way: open source becomes distribution, while the runtime becomes monetization.

The talk supports this point with numbers from Amplifying's study: Claude Code selected shadcn/ui in 64 of 71 responses with an extracted primary pick (90.1%), and Vercel in every extracted deployment response for Next.js and React SPA. The footnote matters. These were 2,430 successful Claude Code responses using three Claude models across four greenfield repositories; Vercel received 86 of 112 deployment picks overall, or 76.8%. This is neither market share nor proof that the selected product is better than its alternatives.

That is why “the agent chose it” does not mean “the market decided.” Selection emerges from the model's training data, its system prompt, the existing stack, documentation quality, the project's visibility in open source, and whether the agent can understand and apply a component locally. In the review of Developer Experience for coding agents, we discussed a standard stack, CLI and API access, fast tests, and intelligible errors. The economic continuation is now visible: agentic ergonomics is becoming not just good DevEx, but a sales channel.

There is a second layer. Rauch expects interfaces and local workflows to be generated for each task more often, while the system of record, data, access permissions, and APIs remain. I would describe this as two software half-lives: a disposable shell and a durable core. The lecture contains its own caveat to the claim that “software is basically free”: for difficult infrastructure code at Vercel, Rauch says, three agents and strong engineers may gather to understand a single line.

Implementation is becoming cheaper, but architecture, verification, security, and system ownership are not necessarily following. In The New SDLC, the harness is the system around a model that makes generation controllable. Here another economic layer appears: the agent and its harness become the point where components are selected.

For a tool vendor, this creates a new question: can an agent discover the product, understand its documentation, install it, recover from errors, and explain why it chose it? For the buyer, the question is different: why did the agent introduce this dependency, and which alternatives did it never consider? DevRel used to compete for a developer's attention. Now it must also compete for a place in the model's world view—and that gatekeeper is often invisible to the user.

#AI #Agents #PlatformEngineering #DevTools #Product #Economics