Vercel has steadily positioned itself at the heart of modern AI software infrastructure. The platform, which enables developers to deploy agents without managing underlying server architecture, now handles roughly 6 million deployments every day. Half of those deployments are initiated by coding agents, and more than 1 trillion tokens pass through the company's AI gateway on a daily basis.

Following the company's ShipNYC conference, Vercel CEO Guillermo Rauch sat down to discuss the current state of AI development, how production realities are reshaping the industry's approach to agents, and why platform companies like Vercel increasingly find themselves in the same competitive space as the major AI laboratories.

From Prototyping to Production

According to Rauch, the shift in developer and enterprise sentiment has been unmistakable. Where the previous year was defined by experimentation - an open-ended phase of building and testing without many constraints - the current moment is defined by a harder question: how do you make agents work reliably at scale?

"Last year was about prototyping," Rauch explained. "The sky's the limit, unleash the agents, everyone can build. We did that, and we learned a lot because we had hundreds of agents organically developed and deployed within the company. Then you started getting into the realities of agents in production, and some of the challenges."

Through that process, Rauch says two clear categories of high-value agent use cases emerged. The first is the coding agent, which he describes as a primary driver of token consumption globally. But as coding agents generate increasing volumes of software, there is a parallel need for infrastructure to host and manage what they produce. The second category is the internal corporate agent - a tool designed to help organizations operate more efficiently by giving employees direct access to their company's data in a conversational, queryable format.

Introducing Eve and Vercel Sandbox

To address the operational and security challenges that come with deploying agents inside companies, Vercel developed two tools. The first is a framework called Eve, which allows teams to define an agent's instructions and capabilities using natural language. The second is Vercel Sandbox, a controlled execution environment designed to give agents the freedom to function intelligently while enforcing strict policies around data access and what information can leave the environment.

Rauch cited data control as one of the most pressing concerns the sandbox addresses. He pointed to a scenario involving coding tools as an illustration of the risk. When a developer installs a tool like Cursor or Devin in the wrong configuration, there is a real possibility that an entire codebase could be transmitted to an external server for training purposes. He recalled a conversation with a senior executive at Airbus as a concrete example - decades of highly specialized aerospace engineering code, written in C++, potentially exposed through a single misconfigured developer tool installation.

"Someone comes in and installs the wrong developer tool and boom, all the code goes out to the cloud for training," Rauch said. "That's a real risk."

The Internal Corporate Agent in Practice

While the coding agent use case has attracted significant attention across the industry, Rauch argues that the internal corporate agent may prove equally transformative over time. He described a specific example from within Vercel's own sales operations to illustrate the concept.

A sales representative focused on growing existing accounts was limited not by her skills or judgment, but by her access to data. To identify which accounts had added the most seats in a given two-week period - a straightforward prioritization task - she would previously have had to wait for a formal engineering project to build a new dashboard. That project might be scheduled for a future quarter, leaving her without the information she needed to act in the present.

With an internal agent in place, that kind of question can be answered immediately. Rauch acknowledged that Vercel itself experienced this bottleneck for years, particularly on the sales and operations side of the business. He described his own unfamiliarity with tools like Salesforce as a limitation that made it difficult to have meaningful influence over parts of the organization outside of engineering.

"Now I feel like I can actually have impact across the entire company," he said, noting that the same Eve framework used for customer-facing agents can also be applied internally to improve productivity.

He also offered a broader observation about what agents are forcing companies to do structurally. In his view, agents require organizations to open up their data and systems through APIs, which runs directly counter to the business model of many established software-as-a-service companies that have historically built competitive advantages by making it difficult for customers to access or export their own data.

"So many of these SaaS giants build their entire kingdoms on trapping your data," Rauch said, "and that's incompatible with agents."

Shifting Relationships with AI Laboratories

Rauch also addressed how enterprise customers are changing the way they think about their relationships with the major AI model providers. A year ago, many organizations were committing to a single lab partner - building everything on top of one provider's models and infrastructure. That approach has largely given way to a more modular mindset.

Customers now understand the full stack - model, harness, data platform, sandbox, gateway - and treat each layer as interchangeable. That means organizations are increasingly mixing and matching models from OpenAI, Anthropic, and Google's Gemini depending on their specific performance and cost requirements.

Rauch noted that Gemini has been growing significantly on Vercel's platform, even though it receives comparatively less coverage in the press. He attributed this to production-focused optimization, where price-to-performance ratios carry more weight than brand recognition or media attention. Open-source and open-weight models are also gaining traction, with DeepSeek and GLM-5.2 both seeing notable adoption increases according to Vercel's internal data.

Competing with the Labs Themselves

The conversation also touched on a growing tension between platform infrastructure companies like Vercel and the AI laboratories whose models they help deploy. When OpenAI recently introduced tools that allow users to publish web content directly from within its own environment, it moved into territory that overlaps with what Vercel offers.

Rauch acknowledged the dynamic but framed it as an opportunity as much as a challenge. He suggested that as more people use tools like ChatGPT to create websites, they will naturally begin asking questions about web hosting - and that those conversations are likely to surface Vercel as a recommended destination. Still, he was candid about the broader competitive reality: as AI platforms expand their capabilities, they will inevitably encroach on the territory of existing infrastructure providers.

The more fundamental question, in his view, is whether the AI model and the agent that uses it will remain tightly coupled - bundled together within a single provider's ecosystem - or whether they will be treated as separate, interchangeable components that developers can mix and match.

Rauch made clear which outcome he is betting on. He drew a parallel to the way software engineering has always worked, with developers pulling in modules, libraries, and building blocks from multiple sources rather than relying on a single monolithic provider. That philosophy, he argued, is what Vercel is bringing to the AI infrastructure market, and it is why the company is actively advocating for open protocols rather than closed, vertically integrated systems.

"We're going to be the AWS of this generation," Rauch said, referencing Amazon Web Services as the model for what a dominant, neutral infrastructure layer looks like. "So obviously we're fighting for a world of open protocols."

The comparison to AWS is deliberate and carries strategic weight. Just as Amazon's cloud division became the foundational layer beneath a generation of internet companies - regardless of what those companies built or which services they used - Rauch appears to be positioning Vercel as the equivalent layer for the AI-native generation of software that is now taking shape.