Deploying AI in Production: How Vercel Separates Agents from Models

To put an AI application into production, companies must separate the model and agent layers. That is the conclusion Vercel CEO Guillermo Rauch draws from the 6 million daily deployments on the platform.

What Vercel's numbers show

Vercel handles 6 million deployments every day, and half of them are initiated by agents rather than humans. More than 1 trillion tokens a day pass through the platform's AI gateway.

These figures are not a forecast but a shift that has already happened: agents have stopped being an experimental tool and have become part of the core infrastructure flow. According to Rauch, two "killer apps" for agents have emerged:

• Coding agents — the main driver of token consumption on the platform
• Enterprise agents — systems that automate internal processes while meeting security and compliance requirements

Why production changes architectural decisions

Rauch names the main trigger for architectural change directly.

"When you optimize for production, you start looking at price/performance."

While an AI product remains a prototype, a developer can work with a single provider for convenience. In a production environment the logic changes: different tasks within an agent workflow call for different ratios of cost to power. One agent step may need deep reasoning, another a fast, cheap call. Tight lock-in to a single model makes optimization impossible.

That is why Vercel's customers now combine providers — OpenAI, Anthropic, Google Gemini and open models such as DeepSeek — based on the characteristics of the task, not out of loyalty to one lab.

This works only with a clean architectural separation: the agent layer (orchestration, tools, memory, prompt logic) exists apart from the model layer. Swapping the model under the hood does not affect the agent logic.

Vercel's new tools for production agents

As of the interview date, Vercel had released two new products:

• Eve — a framework in which agent instructions and skills are described in natural language instead of code
• Vercel Sandbox — an isolated environment that limits agents' access to data while preserving their computing capabilities

Both products aim to make agent infrastructure manageable and predictable in production. Rauch also speaks openly about competitive pressure: the large AI labs are moving into adjacent niches themselves — hosting, deployment, infrastructure. Vercel's answer is to bet on open protocols and position itself as an independent infrastructure layer that does not depend on any single lab.

What this means

With more than a trillion tokens a day, Vercel sees the market's transformation earlier than most analysts: the move from a single-provider strategy to a multi-model architecture is already an observable pattern. Rauch positions Vercel as "the AWS of this generation" in AI infrastructure — a platform on which applications are built on top of any model.

Frequently asked questions

How many tokens does Vercel process per day?

More than 1 trillion tokens a day pass through Vercel's AI gateway. Half of the 6 million daily deployments are launched by agents, not humans.

What is Eve from Vercel?

Eve is a new Vercel framework that lets developers describe agent instructions and skills in natural language instead of code. The tool was introduced in July 2026.

Why did Vercel release Sandbox?

Vercel Sandbox limits agents' access to data while preserving their computing capabilities. It responds to enterprise customers' demands for security and manageability in production.

Why should AI models and agents be separated in production?

Companies that have taken AI applications into production inevitably move to a multi-model architecture with separated layers. According to Vercel, 50% of the 6 million daily deployments are initiated by agents.

How big is the demand for AI applications in production?

Vercel handles 6 million deployments a day, half initiated by agents. More than 1 trillion tokens a day pass through the platform's AI gateway.