Vercel has become an important company in AI software because it sits close to where agents turn ideas into running applications. The company is known for cloud infrastructure that lets developers deploy agents without managing servers, and it now handles 6 million deployments a day.
According to Vercel CEO Guillermo Rauch, half of those deployments are triggered by coding agents. More than 1 trillion tokens also move through the company’s AI gateway daily, showing how much agent activity is already tied to production infrastructure.
From prototypes to production
Rauch described a shift in the AI market after a year dominated by experimentation. Last year, the mood was about prototyping, giving agents broad room to build and seeing what they could do. Vercel took part in that wave internally, with hundreds of agents developed and deployed inside the company.
That experimentation led to a more practical phase. Once agents are used in production, the core questions change. It is no longer enough to ask whether an agent can complete a task. Companies need to understand what the agent can access, what it did, and how its actions can be reviewed afterward.
Rauch pointed to two major agent use cases that are already proving useful. The first is the coding agent. The second is the internal agent that helps run a company.
Those two use cases create different infrastructure needs, but they share one theme: agents need controlled access to tools, data and deployment environments.
Why coding agents create infrastructure demand
Coding agents are one of the clearest drivers of token use. Rauch said they are driving a large amount of token utilization in the world. But when agents produce a lot of software, that software still needs somewhere to go.
That is where a company like Vercel becomes central. If agents can generate, modify or deploy applications, the platform around them has to support frequent deployment and safe execution. The model may write code, but the agent’s work becomes valuable only when it can be shipped, hosted and managed.
This also makes the boundary between AI labs and infrastructure platforms more important. As model providers add more tools, they move closer to the work already handled by deployment platforms. Rauch said it is a natural next step for labs to host small websites, but he also framed that as an opening for Vercel because users may then associate tools like ChatGPT with making websites and ask more questions about web hosting.
The larger question, in Rauch’s view, is whether the model and the agent remain tightly coupled. One path is a system where intelligence, tools and hosting come from one place. The other path is closer to traditional software engineering, where teams combine modules, libraries and building blocks from different providers.
The internal agent as a business tool
Rauch’s second major use case is the internal corporate agent. This kind of agent is not mainly about writing code. It is about helping employees work with company data and tools in ways that previously required dashboards, engineering support or manual processes.
He gave the example of a sales rep at Vercel who works on the install base and is responsible for growing existing accounts. Her bottleneck was not creativity, intelligence or relationship building. It was access to the right data at the right time.
In Rauch’s example, the rep wanted to know which accounts had added the most seats in the last two weeks so she could prioritize her work. Before agents, that kind of question might have required waiting for a new sales dashboard project. With an internal agent, the goal is to make that question directly accessible through APIs and company systems.
Rauch connected this to a broader change in how companies expose their data. He said agents are forcing companies to open up, because agents need access to information and tools to complete work. He also argued that SaaS companies built around trapping customer data are incompatible with agents.
Control, audit trails and sandboxes
The power of internal agents creates a security and governance problem. If an agent can help an employee by reaching into company systems, the company also needs to know what data the agent accessed and what tool calls it made.
Rauch said the challenge is to securely access data, audit what the agent is doing and keep a trail of tool calls and access controls. Vercel’s answer includes a framework called Eve, where agent instructions and skills can be laid out in natural language.
Another piece is Vercel Sandbox. Rauch described it as a way to put an agent in a controlled environment while still letting it express its intelligence. Policy can then be applied to what data the agent can access and what data can leave the sandbox.
The biggest advantage of the sandbox, according to Rauch, is data control. He pointed to the risk that a coding IDE such as Devin or Cursor, if used with the wrong setting, could train on an entire codebase. He described a conversation with the president of Airbus and raised the example of decades of specific C++ code for aerospace engineering leaving for cloud training because someone installed the wrong developer tool.
Model choice becomes a production decision
Rauch also described a change in how companies work with major AI labs. Last year, many customers picked one lab partner and planned to build around OpenAI or Anthropic. Now, he said, more teams understand that the model, harness, data platform, sandbox and gateway can be treated as plug-and-play pieces.
That approach means teams can use OpenAI, Anthropic or Gemini. Rauch said Vercel is seeing a lot of growth from Gemini, even though it is not in the news as much, because production teams look closely at price/performance. He also said open models including DeepSeek and GLM-5.2 are taking off.
The implication is that AI infrastructure may be shaped less by loyalty to one model provider and more by practical assembly. Companies want to choose the model that fits the job, connect it to the right tools, control the data flow and deploy the result safely.
That is the market position Rauch is arguing for. Vercel wants a world of open protocols, where the model is one building block among several rather than the whole agent system. As agents move from prototypes into daily business work, the fight is increasingly about who controls the stack around them.