Vercel, recognized for its cloud platform that empowers developers to deploy agents without the need for server management, has emerged as a pivotal player in the AI software landscape. The company manages approximately 6 million deployments daily, with coding agents responsible for half of these actions, while over 1 trillion tokens traverse its AI gateway each day.
Following the recent ShipNYC conference, we had the opportunity to discuss the current AI landscape with Vercel’s CEO, Guillermo Rauch. In this interview, he shared insights on the evolving role of platform companies like Vercel and their competition with leading laboratories. Below is a summary of our conversation.
This year, there seems to be a renewed sense of purpose in the community, with fewer experimental projects and a stronger emphasis on practical application. How has this shift been reflected at Vercel?
Last year focused on initial explorations. It was about unleashing potential and encouraging widespread development. We gained valuable lessons through the deployment of many agents internally, facing the realities of using agents in a production setting and overcoming various challenges.
One key learning was identifying the most impactful applications of agents. The first is undoubtedly the coding agent, which significantly contributes to global token use. However, as software production increases, it’s essential to have an effective management system. The second major application is the internal agent that assists with management tasks. This raises questions of data security, auditing, and tracking data accessibility that the agent utilizes to perform its functions.
To address these concerns, we developed a framework called Eve, which allows for defining agent instructions and capabilities in natural language. Additionally, we created Vercel Sandbox, designed to restrict an agent’s environment yet still allow it to optimize its performance within set data access policies.
What risks does this approach help mitigate?
The primary advantage of using Sandbox is the ability to control data. A significant risk with AI tools, such as coding IDEs, is that they may inadvertently train on entire codebases in inappropriate settings. For instance, I once spoke with the president of Airbus who emphasized the importance of safeguarding decades of specialized C++ code for aerospace. A mishap with a developer tool could lead to sensitive code being sent to the cloud for training.
Can you elaborate on the second key use case? We’re familiar with coding agents, but what does an internal corporate agent look like?
Imagine a sales representative at Vercel focused on expanding existing accounts. For her, the biggest constraint isn’t creativity or intelligence; it’s access to data. Previously, she might have struggled to identify which accounts were growing quickly, needing to wait for a new sales dashboard project to be completed. Now, using our technology, she can directly access critical sales data, providing her with the necessary insights to prioritize her tasks effectively.
At Vercel, we faced similar challenges for years, especially on the sales engineering side. Although our R&D team advances rapidly, the sales process lagged, leaving me feeling ill-equipped when I first began working with Salesforce.
Today, I can positively impact the entire company, as Eve can support both customer-oriented agents and enhance productivity through API integration. This accessibility encourages organizations to become more transparent, which has significant long-term implications. Many software giants are built on restrictive data practices, but agents demand a different approach.
How do you foresee changes in client relationships with major AI labs?
Last year, many companies gravitated towards selecting a single lab partner, committing to build everything solely on OpenAI or Anthropic. Now, however, they recognize the modular nature of AI components. They can combine various platforms—be it OpenAI, Anthropic, or Gemini. We’ve noticed Gemini’s growth, even if it’s not highlighted in the news, as businesses seek production-optimized options. When considering production needs, price-performance ratios have become paramount, and Gemini offers great characteristics in that regard. Open models like Deepseek and GLM-5.2 are also gaining traction.
You mentioned direct competition with AI labs; how does this manifest?
The recent introduction of tools by OpenAI that allow users to publish directly online is a logical progression for them. This opens avenues for viewers to see ChatGPT as a website creation tool, leading to potential recommendations for our services. As labs expand their capabilities, they inevitably compete with established infrastructure platforms.
We are now at a point of deciding whether to couple models with agents. Should all intelligence be sourced from a single provider, or should we adopt a modular approach where components come from various sources? The latter aligns more closely with traditional software engineering principles, which we aim to promote. Our goal is to position ourselves as the leading infrastructure provider of this generation, advocating for open protocols.
