Vercel CEO Guillermo Rauch Discusses the Challenge of Distinguishing Models from Agents

Vercel has established itself as a leader in cloud infrastructure, allowing developers to deploy agents seamlessly without server management. It has consistently emerged as a significant player in AI software, managing 6 million deployments each day, half of which are initiated by coding agents, alongside processing over 1 trillion tokens through its AI gateway daily.

After the recent ShipNYC conference, we spoke with Vercel CEO Guillermo Rauch about the current AI landscape and how platform companies like Vercel are standing up against larger laboratories. Here’s a lightly edited transcript of our discussion.

There appears to be a fresh momentum in the community this year, shifting from pilot projects to tangible applications. Have you noticed this shift among your clients, and what has it looked like at Vercel?

Last year was focused on prototyping. The mantra was to unleash agents, empowering everyone to create without restriction. We embraced this, gathering invaluable insights from hundreds of organically developed agents deployed internally. However, we soon encountered the challenges of deploying agents in production.

A standout lesson for me highlighted two primary use cases for agents. The first is the coding agent, which plays a significant role in global token utilization. However, generating extensive software demands a repository. The second notable application involves internal agents that enhance company operations. The main challenge is ensuring secure data access, monitoring agent activities, and establishing a clear trail of tool calls and access controls necessary for various tasks.

To tackle these challenges, we created a framework called Eve, where you can define an agent’s tasks and capabilities in natural language. We also introduced Vercel Sandbox, providing a controlled environment for agents to operate. This enables them to exercise their intelligence while adhering to data access policies and outbound data flow regulations.

What kinds of challenges does this help you avoid?

The sandbox significantly improves data control. A notable risk associated with AI that I often ponder is the use of coding IDEs like Devin or Cursor. If misconfigured, they could unintentionally train on your entire codebase. I remember discussing this concern with the president of Airbus, who has a vast repository of specialized C++ code for aerospace engineering. One wrong developer tool setup could risk having that critical code uploaded to the cloud for training.

I’m curious about that second vital use case. While coding agents are familiar, what does an internal corporate agent entail in practice?

Imagine a sales representative at Vercel focused on nurturing existing accounts. Her creativity, intelligence, and relationship-building abilities aren’t the issue; it’s data access. In the past, she could not inquire about which accounts were experiencing the fastest growth; she had to wait for a Q1 project to finalize a new sales dashboard.

We faced that limitation for years at Vercel, which was frustrating, especially given our strong agility on the R&D side. Yet, on the sales side and with Salesforce engineering, I initially felt out of my element as I had never navigated Salesforce prior to this.

Now, I can genuinely influence the entire company. Eve can enhance both our customer-facing agents and overall productivity. The technology remains focused on APIs, encouraging companies to embrace transparency, leading to significant long-term implications. Many SaaS giants thrive on data entrapment, which is at odds with the principles driving agents.

How do you envision client relationships evolving with major AI labs?

Last year, many firms aligned with a single lab partner—committing to build everything on OpenAI or Anthropic. Now, they’re realizing the technology’s modular nature—model, harness, data platform, sandbox, gateway—where each component can be swapped out. You might choose OpenAI, Anthropic, or Gemini. We’re seeing substantial growth in Gemini, even if it’s less publicized, as companies increasingly emphasize production. Ultimately, with production optimization, the focus is on cost-effectiveness, and Gemini models provide impressive price/performance ratios. Open models like DeepSeek and GLM-5.2 are also gaining traction, as the data indicates.

In some respects, it seems you’re in direct competition with these labs. Recently, OpenAI introduced new tools that facilitate direct web publishing within their ecosystem.

It’s a natural evolution for them to create mini websites. This offers us a great opportunity, as users may start seeing ChatGPT as a website creation tool. If users continue to ask about web hosting, the model could suggest our services. However, as platforms roll out more features, they inevitably begin to compete directly with established infrastructure providers.

We now face a decision regarding the integration of models and agents. Will all intelligence originate from a single source, or will a modular approach permit us to draw components from various providers? This modularity echoes traditional software engineering, which is exactly what we strive to offer. We aim to be the AWS of this era, promoting a framework of open protocols.