Vercel CEO Guillermo Rauch on the fight to split off models from agents

Vercel, a key player in cloud infrastructure, has positioned itself at the forefront of AI software development. With an astonishing 6 million daily deployments—half initiated by coding agents—and over 1 trillion tokens processed through its AI gateway each day, the company is making waves in the tech world. Following Vercel’s recent ShipNYC conference, CEO Guillermo Rauch shared insights into the evolving landscape of AI and the competitive dynamics between platform companies and major research labs. This year, the atmosphere in the AI community feels markedly different, according to Rauch. He observes a shift towards practical implementation, moving away from experimental pilot programs. Vercel's journey reflects this trend; last year focused on prototyping and unleashing coding agents, leading to valuable lessons learned from deploying hundreds of agents internally. The realities of operating agents in a production environment revealed both opportunities and challenges. Rauch identifies two primary applications for agents: coding agents, which significantly contribute to token usage, and internal agents that streamline company operations. However, integrating these agents securely poses questions about data access and auditing. To address these issues, Vercel has developed frameworks like Eve, which allows natural language instructions for agents, and Vercel Sandbox, a controlled environment that ensures data governance while enabling agent functionality. A critical advantage of the Sandbox is its data control capabilities. Rauch warns of the potential risks associated with coding tools that might inadvertently expose sensitive codebases, highlighting the importance of safeguarding proprietary information. He recalls a conversation with the president of Airbus, emphasizing how a single misstep with developer tools could compromise decades of valuable code. Rauch further elaborates on the internal agent use case, illustrated by a sales representative at Vercel. Traditionally, her productivity was hampered by a lack of accessible data, preventing her from identifying key accounts. With the implementation of Eve, she can now leverage real-time data to optimize her efforts, demonstrating how agents can enhance overall company efficiency. As the landscape shifts, Rauch notes that client relationships with AI labs are evolving. Rather than committing to a single provider, companies now recognize the modularity of AI components, allowing for a flexible approach in utilizing various models and platforms. The emergence of Gemini, for instance, highlights the growing focus on optimizing production based on performance and cost. Vercel's competitive stance is further illustrated by its response to new tools released by OpenAI, which could potentially position them as rivals. The conversation around whether to couple models and agents or to adopt a modular approach is critical, as companies seek to build a robust infrastructure for the future of AI. Rauch envisions Vercel as a transformative force in this landscape, aspiring to be the AWS of the new generation and advocating for open protocols in AI development.