Sourcery: Guillermo Rauch on Building the Internet for AI Agents

SPEAKER_1 (0:01)

When software starts making choices on its own, the internet changes shape. In this episode of Sourcery, we're diving into a conversation with Vercel CEO Guillermo Rauch about coding agents and what happens when machines increasingly talk to machines. And the numbers really frame this moment. Vercel just raised $300 million in a Series F round, hit a $9.3 billion valuation, and reported about $600 million in annualized revenue. But here's what caught my attention—Rauch says agentic deployments jumped from 3% at the start of the year to 60% now. That's a massive shift in just months. They added 1.5 billion deployments in 10 months alone.

So what's driving that growth?

Rauch roots it in Vercel's culture—values like 'iterate to greatness' and 'dig deep.' He says that spirit comes from building tools in public, staying open source, and treating the company itself as a product. That openness helped turn Next.js into a foundation for modern web development. Right, and the big theme here is this move from human-centric software to agentic software.

Rauch points out that AI has lowered the barrier to shipping code while massively increasing the volume. But he's careful to call code a liability that must be handled carefully.

Which is why he emphasizes what he calls 'rejecting non-understanding.' Humans need to read and verify what AI produces. That concern extends to the internet itself—Rauch argues the web is evolving from pages designed for human eyes into systems that serve agents directly. So Vercel is building what they call 'agentic developer experiences,' supporting protocols like MCP so machines can discover tools, deploy software, and transact more effectively.

What about the model market? It's changing fast. Vercel acts as a broad token distribution layer, and Rauch says 80% of the tokens on their network now go to open models. Enterprises want optionality, not lock-in. They want to swap models as easily as changing a line of code. But security becomes the pressure point. Rauch warns about prompt injection, autonomous agent behavior, and the need for sandboxes or 'jails' that keep systems contained. They recently ran a sandbox challenge and actually helped uncover a serious Linux vulnerability. That shows how quickly offensive and defensive capabilities are escalating. And the hiring philosophy follows the same logic—Rauch says proof of work matters more than a static resume. He wants people who can command agents, communicate clearly, and work close to the customer and the code.

He even describes Vercel migrating a Mercedes stack in minutes with an agent. But through all of it, Rauch returns to reliability and speed. AI can generate code quickly, but businesses still need durable systems, strong security, and long-term maintenance. Exactly. The real shift isn't about writing every line by hand anymore—it's about making better decisions about where to apply token power. And Rauch sounds both pragmatic and optimistic about it. His final message? 'Everything is going to be okay.'