The Shift from Prototyping to Production

Last year, the AI industry was defined by the "sky’s the limit" phase. Developers were rushing to build agents, companies were running pilot programs, and the hype cycle was in full swing. But according to Vercel CEO Guillermo Rauch, that era has ended. The focus has shifted from what agents can do to how they can actually function in a secure, high-stakes production environment.

"Last year was about prototyping," Rauch said following Vercel’s ShipNYC conference. "We learned a lot because we had hundreds of agents organically developed and deployed within the company. Then you start getting into the realities of agents in production."

Today, Vercel sees 6 million deployments a day, with half triggered by coding agents. More than 1 trillion tokens flow through the company’s AI gateway daily. This massive volume has given Rauch a front-row seat to the friction points that emerge when AI moves from a playground to a corporate engine.

The Two Killer Apps of Agents

While the industry remains obsessed with the capabilities of frontier models, Rauch argues that the real value is crystallizing around two specific use cases. The first is the coding agent—the engine behind the current surge in token utilization. The second, and perhaps more transformative, is the internal corporate agent.

"The challenge there is, how do you securely access data? How do you audit what the agent is doing?" Rauch explained. "How do you get a trail of all the tool calls and access controls that the agent had to incur in order to get a job done?"

To address this, Vercel has introduced "Eve," a framework designed to define agent instructions and skills in natural language, and "Vercel Sandbox," a secure environment that cages agents. This allows them to express intelligence while applying strict policies on data access and egress. For Rauch, this is a matter of corporate survival. He points to the risk of sensitive intellectual property—like the decades of aerospace code held by companies like Airbus—being inadvertently ingested by a coding IDE for model training.

Why SaaS Giants Are Losing Their Grip

Perhaps the most significant trend Rauch observes is the "unbundling" of the AI stack. Last year, companies were content to hitch their wagons to a single lab partner, building their entire strategy around OpenAI or Anthropic. That era of vendor lock-in is fading.

"Now they’re saying, 'I understand how this all works—model, harness, data platform, sandbox, gateway—every piece is plug and play,'" Rauch said. "You can use OpenAI, you can use Anthropic, or you can use Gemini."

This modularity is forcing a reckoning for legacy SaaS companies. Many of these firms built their business models on "trapping" customer data, a strategy Rauch views as fundamentally incompatible with the agentic future. Agents require open access to data to be useful, and companies that refuse to provide that access will find themselves bypassed by more agile, agent-first competitors.

What This Means for Developers

For developers, the message is clear: stop treating the model as the entire product. The model is a commodity, and price-performance is becoming the primary metric for production-grade systems. As developers look to optimize costs, they are increasingly turning to a mix of providers, including open-source models like DeepSeek and GLM-5.2.

"The data doesn’t lie," Rauch noted. "When you’re optimizing for production, you start looking at price/performance, and Gemini models have awesome price/performance characteristics."

Key Takeaways

  • Infrastructure over Hype: The industry has moved past the "prototyping" phase, with companies now prioritizing security, auditing, and data control for AI agents in production.
  • The Unbundling of the AI Stack: Businesses are moving away from single-lab dependencies, treating models, gateways, and sandboxes as modular, plug-and-play components.
  • Data Control is the New Frontier: Tools like Vercel Sandbox are becoming essential to prevent sensitive corporate data from being leaked or used for model training by third-party tools.

The Next Decision Point

As Vercel and other infrastructure providers continue to build out these layers, the competition with the labs themselves will only intensify. When OpenAI releases tools that allow users to publish directly to the web, they are encroaching on the territory of deployment platforms. The question for the coming year isn't just which model is smartest, but which platform can provide the most secure, transparent, and cost-effective "cage" for the agents that will eventually run the modern enterprise.