Vercel's Guillermo Rauch on Why AI Models and Agents Should Stay Separate

Vercel has quietly become one of the most important companies in AI software, powering cloud infrastructure that lets developers deploy agents without managing their own servers. According to the company, it now handles 6 million deployments a day, half of them triggered by coding agents, while more than 1 trillion tokens pass through its AI gateway daily.

Following the company's ShipNYC conference, CEO Guillermo Rauch spoke about the current moment in AI, the practical realities of running agents in production, and how platform companies increasingly find themselves competing with major AI labs.

From prototyping to production

Rauch said the mood in the developer community has shifted from experimentation toward making systems work reliably in practice. Last year, he explained, was about prototyping and letting teams build freely. Vercel itself developed and deployed hundreds of agents organically, an experience that surfaced the challenges of running them at scale.

The biggest lesson, he said, was identifying the two standout use cases for agents. The first is the coding agent, which drives much of the world's token usage. But producing large volumes of software creates a follow-on need: somewhere to deploy and host it. The second, according to Rauch, is the internal agent that helps run a company.

That second category raises harder questions around security. Companies need to control how agents access data, audit their actions, and keep a trail of every tool call and access control involved in completing a task.

Building guardrails with Eve and Sandbox

To address those concerns, Vercel created a framework called Eve, which lets teams define an agent's instructions and skills in natural language. A second tool, Vercel Sandbox, effectively places an agent in a controlled environment where it retains freedom to operate but is subject to policies governing which data it can access and what information can leave the sandbox.

The main advantage, Rauch said, is data control. He pointed to the risk of coding tools training on a company's entire codebase if configured incorrectly. He recalled a conversation about aerospace engineering, where decades of highly specialized C++ code could be exposed to the cloud for training if the wrong developer tool were installed.

What an internal corporate agent looks like

Rauch illustrated the internal use case with an example from Vercel's own sales team. A representative focused on growing existing accounts had long been limited not by skill or creativity but by access to data. She wanted to know which accounts were expanding fastest, or which five had added the most seats in the past two weeks, so she could prioritize her work.

Previously, answering that kind of question meant waiting for a dedicated sales dashboard project to be completed. Rauch described that bottleneck as a source of frustration for years, contrasting Vercel's fast-moving research and development with slower progress on internal sales systems.

Now, he said, the same underlying technology powers both customer-facing agents and internal productivity tools, because it all runs on APIs. In his view, agents are pushing companies to open up their data, with significant long-term consequences. Many large software-as-a-service providers, he argued, built their businesses on locking in customer data, an approach he sees as incompatible with how agents work.

Changing relationships with AI labs

Rauch also described a shift in how clients work with major AI labs. A year ago, many chose a single partner and committed to building everything on one provider. Now, he said, customers treat every component — model, harness, data platform, sandbox, and gateway — as interchangeable, mixing offerings from different providers.

He noted growing adoption of Gemini, attributing it to strong price-to-performance characteristics as teams optimize for production. He also cited rising interest in open models, saying the underlying usage data reflects those trends.

At the same time, Rauch acknowledged that platforms and labs increasingly compete directly. He referenced a recent release that lets users publish directly to the web from within an AI lab's environment. He framed that development as both a natural step for the labs and an opportunity, since users who begin thinking of a chatbot as a website-building tool may then ask it about web hosting and receive recommendations.

The fight over coupling models and agents

For Rauch, the central question now is whether the model and the agent will be coupled together. He asked whether users should get all their intelligence from a single source, or instead assemble building blocks from different providers — an approach he compared to traditional software engineering.

He positioned Vercel as bringing that modular philosophy to market, describing an ambition to become "the AWS of this generation" and advocating for a world built on open protocols.

As the line between AI models, agents, and infrastructure continues to blur, the debate Rauch describes could shape how the next wave of AI software is built. Share this article to keep the conversation going.

Source: TechCrunch