Vercel, recognized for its cloud infrastructure enabling developers to deploy agents without server management, has discreetly established itself as a pivotal player in AI software. The company observes 6 million deployments daily, with half initiated by coding agents, and over 1 trillion tokens processed through its AI gateway each day.

Following the company’s ShipNYC conference last week, we engaged in a discussion with Vercel CEO Guillermo Rauch about the current AI landscape and how platform companies like Vercel find themselves in competition with major research laboratories. Below is a lightly edited transcript.

This year seems to bring a fresh energy in the community, with fewer pilot projects and a greater emphasis on practical implementation. I suspect you’ve encountered that with clients, but I’m interested in how that journey has unfolded internally at Vercel.

Last year focused on prototyping. The potential was limitless, release the agents, everyone can create, and so forth. We accomplished that, gaining valuable insights from having numerous agents organically created and deployed within the organization, only to later confront the actualities of agents in production and their associated challenges.

The most significant takeaway for me was identifying the major use cases, the two standout applications of agents. The first is, of course, the coding agent. This is a key driver of token usage globally, but with such a high volume of software production, there must be a repository. The second major application of agents is the internal agent that facilitates company operations. The main issue there is securing data access, auditing agent activities, and establishing a history of all tool interactions and access permissions that the agent had to navigate for task completion.

To address this, we developed a framework named Eve, which allows you to outline an agent’s directives and abilities in natural language. Another tool is Vercel Sandbox, which confines the agent in a controlled environment. While it can still express its intelligence, we can enforce policies regarding what data it can access and what data can exit the sandbox.

What kinds of issues does this help you avoid?

The biggest benefit of the sandbox is data management. A significant risk of AI that I consistently consider is when you utilize a coding IDE like Devin or Cursor; if you’re in an inappropriate environment, they might train on your entire codebase. I remember discussing this with the president of Airbus. They possess decades of highly specialized C++ code for aerospace engineering. If someone mistakenly installs the wrong developer tool, all the code can be sent to the cloud for training.

I’m interested in learning more about that second primary use case. We are familiar with coding agents, but what does an internal corporate agent look like in reality?

For example, there’s a sales representative at Vercel. Her role is to expand existing accounts. The bottleneck for individuals like her hasn’t been a lack of creativity, intelligence, or relationship-building skills; it’s been data. “I need to know which accounts are growing rapidly. Provide me with the five accounts that have increased the most seats in the past fortnight so I can prioritize.” In the past, she couldn’t ask that question and had to wait for a Q1 project for a new sales dashboard to conclude.

We faced that bottleneck for years at Vercel, which was quite frustrating because, on the R&D side, we are the most agile company globally. However, on the sales front, the Salesforce engineering aspect was something I was utterly unprepared for. I had never used Salesforce before I started.

Now, I believe I can truly impact the entire organization because Eve can be utilized for our customer-serving agents and enhance productivity. The same technology simply utilizes APIs. Agents are compelling businesses to become more open, leading to significant long-term consequences. Many of these SaaS giants have constructed their empires by entraping your data, which is incompatible with agents.

How do you perceive client relationships with the large AI labs evolving?

Last year, many were selecting a singular lab partner, committing to building everything on OpenAI or Anthropic. Now, they’re realizing how it all fits together — model, harness, data platform, sandbox, gateway — everything is modular. You can employ OpenAI, Anthropic, or Gemini. We are witnessing significant growth in Gemini, even if it isn’t prominently featured in the news, because companies are now prioritizing production. The truth is, when optimizing for production, you start considering price/performance, and Gemini models exhibit excellent price/performance metrics. Open models are also gaining traction, with DeepSeek and GLM-5.2 becoming increasingly popular. The data speaks for itself.

There are areas where you are in direct competition with the labs as well, correct? Just recently, OpenAI unveiled a new suite of tools that enables direct web publishing without leaving the OpenAI ecosystem.

It’s a logical progression for them to host small websites. It creates a fantastic opportunity for us because individuals will begin to view ChatGPT as a platform for website creation. If they persist in querying the model about web hosting, it may recommend our services. However, you are correct that as models and platforms enhance their capabilities, they directly compete with the existing infrastructure platforms.

I genuinely believe we are at a crossroads regarding whether the model and the agent will be interconnected.

Will you derive all your intelligence from one source? Or will you receive a module, library, or building block from one provider and then build upon it? This aligns more with traditional software engineering, which is precisely what we’re introducing to the market. We aspire to be the AWS of this era, thus we are striving for a landscape of open protocols.