Vercel's split-models-from-agents strategy reshapes production AI playbooks

The next major shift in AI infrastructure is the move to separate models from the agents that consume them, enabling open, modular stacks and multi-provider deployments for production AI. Vercel positions itself as an infrastructure platform aiming to be an "AWS of this generation" by emphasizing modular APIs, policy-enforced sandboxes, and governance for production agents.

What happened

After Vercel's ShipNYC conference, CEO Guillermo Rauch shared insights in a TechCrunch interview on how the company has become a central player in AI software. Vercel currently sees 6 million deployments a day, about half triggered by coding agents, and more than 1 trillion tokens flow through its AI gateway daily.

Rauch described the two "killer apps" of agents: coding agents that drive software production and internal agents that help run the company. The critical challenge with internal agents is securely accessing data, auditing actions, and maintaining a trail of tool calls and access controls.

To solve this, Vercel introduced two tools:

• Eve framework: Allows developers to define an agent's instructions and skills in natural language.
• Vercel Sandbox: Puts the agent in a "cage" where it retains its intelligence but has policy applied on what data it can access and what data can leave the sandbox.

Rauch cited a concrete example: a sales rep working on install base growth who could not ask "give me the five accounts that have added the most seats in the last two weeks" without waiting for a Q1 sales dashboard project. Using Eve, her internal agent now extracts those signals from live CRM data directly.

He also noted that companies are moving away from picking a single lab partner. They now use OpenAI, Anthropic, or Gemini as plug-and-play components, with Gemini seeing significant growth due to price/performance. Open models like DeepSeek and GLM-5.2 are also gaining traction.

Why AI builders should care

Pricing and performance are central to production deployments as teams mix models from multiple providers and open-source options. Rauch emphasized that when optimizing for production, teams look at price/performance, and Gemini models have strong characteristics. The data "doesn't lie" that multi-provider stacks are taking off.

Platform-level modularity and data governance become competitive differentiators as firms shift from prototyping to production-ready AI agents. The debate between coupling models to a single provider versus assembling modular components mirrors traditional software engineering practices.

Practical implications

For AI builders deploying internal agents, several considerations emerge:

• Governance tooling: Evaluate sandbox environments (like Vercel Sandbox) and data access policies to prevent data leakage. Rauch warned about risks like a coding IDE training on an entire codebase, citing a conversation with Airbus about proprietary C++ code being sent to the cloud for training.

• Modular model strategy: Adopting a multi-provider model stack reduces vendor lock-in and allows teams to tailor model choices to tasks, improving production reliability and cost control.

• Audit trails: Platforms that offer clear policy enforcement and audit trails for agent actions will be favored in production environments.

Caveats

The narrative is based on an interview with Vercel's CEO and the company's public claims. Specific deployment metrics, model performance numbers, and governance outcomes are not independently verified in this pack. Vendor claims about future open protocols and market dynamics may evolve as the AI tooling ecosystem shifts. Pricing and performance can vary by workload and provider.

FAQs

What does it mean to split off models from agents in AI systems?

It refers to decoupling the machine learning model from the agent that consumes it, enabling a modular approach where different models can be swapped or combined with agents. This aims to improve production flexibility and governance by avoiding tight coupling of data, tooling, and models. As Vercel CEO Guillermo Rauch explains, this is about deciding whether to get all intelligence from one place or assemble building blocks from multiple providers.

Why is production readiness a focus when separating models from agents?

Production readiness emphasizes reliability, data governance, access controls, and audit trails as agents operate on live data and systems. The move to production-ready agents shifts concerns from prototyping to scalable, auditable workflows. Rauch noted that last year was about prototyping and unleashing agents, but this year teams are optimizing for production, looking at price/performance and governance.

How do tools like Vercel's Eve framework and Sandbox help with agent safety and policy enforcement?

Eve helps define an agent's instructions and capabilities in natural language, while Vercel Sandbox cages the agent and enforces data-access and data-exit policies. Together they aim to reduce data leakage and improve auditable behavior of production agents. Rauch described the Sandbox as a place where an agent can have freedom to express its intelligence but policy is applied on what data it can access and leave the sandbox.

What role does platform API modularity play in choosing model providers for AI agents?

Modularity allows teams to mix and match models from multiple providers (OpenAI, Anthropic, Gemini) and open models, enabling price/performance optimization. This mirrors software engineering practices of assembling building blocks rather than tying to a single vendor. Rauch observed that companies are now treating each piece (model, harness, data platform, sandbox, gateway) as plug and play, and Gemini is gaining adoption for its price/performance.