Vercel CEO Guillermo Rauch has called for a structural separation between AI models and the agents that use them, warning that developer tools training on proprietary codebases pose an existential data risk. Speaking to TechCrunch after the company's ShipNYC conference, Rauch revealed Vercel now handles 6 million deployments daily — half triggered by coding agents — with more than 1 trillion tokens flowing through its AI gateway.
Vercel's new Eve framework lays out agent instructions in natural language, while Vercel Sandbox puts agents "in a little cage" with policy controls on data access and egress. Rauch cited a conversation with the president of Airbus, who feared a single developer installing the wrong IDE could exfiltrate decades of proprietary aerospace code.
Analysis
Rauch's model-agent split framing is strategically convenient — Vercel profits from being the middleware — but the argument is sound. As coding agents proliferate, the boundary between development environment and training pipeline blurs. Platform companies that control the deployment layer gain enormous leverage over which models get used and how.
