Vercel CEO Guillermo Rauch on the Crucial Split Between Models and Agents
As the AI industry transitions from experimental prototyping to production-ready applications, the infrastructure powering these tools is undergoing a massive shift. Vercel CEO Guillermo Rauch argues that the future of AI lies in decoupling the underlying models from the agents that execute tasks, moving toward a modular ecosystem.
The Shift from Prototyping to Production
Last year, the AI landscape was defined by "unleashing agents" and testing the limits of what was possible through pilot programs. However, Vercel’s recent data indicates a significant maturation in how developers approach AI. The company currently manages 6 million deployments daily, with more than 50% of those triggered by coding agents. Furthermore, over 1 trillion tokens flow through Vercel’s AI gateway every single day.
According to Rauch, the industry has moved past the "sky's the limit" phase and is now grappling with the realities of production: security, data access, and auditability. This shift has highlighted two primary "killer apps" for agents: coding assistants and internal corporate agents designed to automate business workflows.
Solving the Security Dilemma: Sandboxes and Data Control
One of the most significant hurdles in deploying autonomous agents is the risk of data leakage. Rauch highlights a critical concern for enterprises like Airbus: if a developer uses an AI-integrated IDE like Cursor or Devin without proper safeguards, proprietary code could inadvertently be used to train public models.
To mitigate these risks, Vercel is developing specialized infrastructure. Their Eve framework allows users to define an agent’s instructions and skills using natural language, while Vercel Sandbox provides a "cage" for the agent. This sandbox allows the agent to exercise intelligence while applying strict policies on what data can be accessed and, more importantly, what data is permitted to leave the environment.
The End of SaaS Data Silos
The rise of agents is also creating a fundamental conflict with traditional SaaS business models. Many software giants have built their value by "trapping" user data within proprietary ecosystems. Rauch notes that this model is inherently incompatible with agents, which require seamless, API-driven access to data to be effective.
Whether it is a sales representative using an agent to query complex Salesforce data or an R&D team automating workflows, agents demand an open architecture. This shift is forcing companies to prioritize data accessibility, which will have long-term implications for how enterprise software is built and sold.
The Fight for an Open AI Infrastructure
As major AI labs like OpenAI expand their capabilities—such as releasing tools that publish directly to the web—platform companies like Vercel face direct competition. Rauch views this as a battle over whether the "model" and the "agent" will be coupled or decoupled.
Vercel is positioning itself as the "AWS of this generation," advocating for a modular approach where developers can swap components like LEGO bricks. Instead of being locked into a single provider, developers are increasingly optimizing for price and performance by mixing models from OpenAI, Anthropic, Google's Gemini, and open-source models like DeepSeek.
Key Takeaways
- Production over Prototyping: The AI industry is moving from experimental pilots to high-scale production, evidenced by Vercel's 1 trillion daily tokens.
- Security is Paramount: Tools like Vercel Sandbox are essential to prevent proprietary data from being leaked into model training sets.
- Modular Architecture: The future of AI software relies on decoupling models from agents, allowing developers to optimize for cost and performance using a multi-model approach.