Agentic AI
Vercel's Rauch: The Future of AI Hinges on Decoupling Models From Agent Platforms
As coding agents drive massive token consumption through its infrastructure, Vercel's CEO argues that platform independence—not lock-in to single AI labs—will define the next wave of AI adoption.
TechCrunch — AI
Vercel has become a quiet powerhouse in the AI infrastructure space, now processing over a trillion tokens daily through its gateway while handling 6 million deployments a day—roughly half driven by AI agents. In a recent conversation, CEO Guillermo Rauch shared his perspective on how the company has shifted from exploration to solving real production challenges.
From Prototyping to Production Reality#
The energy in AI has fundamentally changed from a year ago, Rauch explained. Early 2025 was dominated by experimentation and proof-of-concept work, but companies are now wrestling with how to actually deploy agents reliably at scale. Within Vercel itself, hundreds of agents were built and deployed, revealing both the potential and the operational friction of real-world agent systems.
Two use cases emerged as genuinely transformative. The first—coding agents—has become obvious to everyone watching AI's impact on software development. The second, less discussed but equally important, is the internal corporate agent that helps run day-to-day business operations. This second category introduced an urgent design challenge: how do you let an agent operate intelligently while maintaining data security, auditability, and access controls?
Two New Tools for Production Safety#
To address these concerns, Vercel introduced Eve, a framework for agents that lets developers define instructions and capabilities in natural language, and Vercel Sandbox, which isolates agent execution while still allowing intelligent decision-making. The sandbox approach prevents one of AI's biggest risks: rogue tools silently exfiltrating proprietary code for training. Rauch cited a conversation with Airbus's leadership as a cautionary example—decades of specialized aerospace engineering code could vanish if a developer unknowingly installs the wrong IDE.
For internal agents, Rauch described a concrete use case: a sales representative drowning in data fragmentation, unable to ask basic questions like "which accounts grew fastest this month?" without waiting months for an engineering project to build a dashboard. Agents can now connect directly to business systems, letting employees ask natural-language questions and get instant answers—but only if the infrastructure enforces strict data boundaries.
