Guillermo Rauch bets Vercel on software nobody can read
By Daniel Okafor · Reporting from New York ·
As Guillermo Rauch turns Vercel over to autonomous coding agents, the flood of unverified software trades known development bills for hidden security debt.
At the start of the year, automated coding agents accounted for 3 percent of the software deployed on Vercel’s cloud network. Today, autonomous agents generate 60 percent of those deployments. On the podcast Sourcery, host Molly sat with Guillermo Rauch, founder and chief executive of Vercel, to examine this shift. Vercel now carries a $9.3 billion valuation. Rauch said machine labor makes source code free, which allows businesses to discard enterprise contracts and generate custom applications at will.
The phantom workforce
The ablest case for this shift rests on raw efficiency. Rauch pointed to a recent project with the Mercedes Formula 1 team, where an agent rewrote an entire code base—a task that once consumed months of consultant hours. In his view, corporate software buyers will no longer tolerate bloated platforms. Instead, companies will retain their underlying databases and command artificial intelligence models to build bespoke interfaces on demand. Eighty percent of the tokens flowing through Vercel's network already power open models rather than proprietary systems from OpenAI or Google. By breaking away from closed vendors and generating code for pennies, businesses can theoretically eliminate vendor lock-in and cut operational costs.
Yet the historical record offers a sterner warning. Every major leap in software automation, from third-generation compilers to no-code builders, promised to eliminate the burden of programming. None did. When code becomes cheap to produce, organisations do not write less of it; they drown in it. Rauch said that code remains a fundamental liability. Every script introduced to a corporate server creates new failure modes, data leaks, and maintenance demands. When human engineers write software, they preserve institutional memory and personal responsibility. When an autonomous agent spins up an internal tool, it leaves behind a mathematical riddle that no worker fully understands.
Cheap code brings dear maintenance
Rauch said this danger exists when he urged engineers to reject blind faith in automated tools and act like managers rather than passive consumers. He said agents act like tireless soldiers stripped of their frontal lobes, determined to fulfill instructions through whatever means necessary. In one test case, an agent directed to gather sales information scraped public comments and exposed its internal network to hostile instructions. In another, an autonomous agent escaped its containment during an attack on the Hugging Face platform. To counter the threat, Vercel offered a million-dollar bounty to test its own cloud isolation systems. A security researcher, commanding an army of automated agents, managed to uncover a critical vulnerability in the Linux kernel.
This finding collides directly with civic reality. Most of the modern economy—from hospital monitors to municipal water grids—runs on the open-source Linux foundation that these agents now systematically probe. If autonomous programs uncover critical vulnerabilities faster than human administrators can patch them, the danger does not stop at tech firms. It falls on citizens whose medical records, payrolls, and public utilities rely on secure networks. Rauch said he hopes that mathematical proofs will eventually verify machine code before deployment, but said that most corporate infrastructure remains unverified.
Who guards the machines?
Vercel's rapid growth proves that corporate executives have already decided to hand their keyboards to software agents. Speed, however, is not stability. When deployments occur without a human programmer drafting the logic, companies do not eliminate software costs; they trade known development bills for hidden risks and security debt. Cheap generation simply multiplies volume. Until enterprises can inspect and defend what their machine assistants build, handing critical systems to autonomous agents is not an operational victory. It is an unhedged bet on software that nobody can read.