Vercel CEO argues for decoupling AI models from agents to improve production workflows

Vercel CEO Guillermo Rauch describes a shift in the AI community from rapid prototyping to production optimization, driven by the deployment of hundreds of agents within Vercel and its customers. He highlights two "killer apps" for agents: coding assistants that generate software and internal corporate agents that surface data for decision-making. These use cases revealed challenges in production, including secure data access, auditability, and tool-call tracing.

To address these challenges, Rauch says Vercel developed a framework called Eve, which allows teams to specify agent instructions and skills in natural language, and a tool called Vercel Sandbox, which constrains agent behavior and enforces data access policies. The sandbox approach aims to mitigate risks such as accidental codebase exposure to cloud-based training systems, a concern Rauch raised with the president of Airbus.

Rauch reports that Vercel now processes more than 1 trillion tokens daily and handles 6 million deployments, half of which are triggered by coding agents. He notes that enterprises are increasingly optimizing for price/performance, leading to broader adoption of models from multiple providers, including OpenAI, Anthropic, Google’s Gemini, DeepSeek, and GLM-5.2.

He argues that the industry is at a crossroads between coupled model-agent architectures and modular, plug-and-play designs. Vercel’s platform enables companies to mix and match models and agents, positioning itself as an infrastructure layer that competes directly with major AI labs. Rauch compares Vercel’s role to AWS, emphasizing the importance of open protocols and modularity in the next generation of AI infrastructure.