Vercel's CEO, Guillermo Rauch, is picking a fight worth understanding. He wants the model and the agent kept separate, so a company can choose each on its own merits instead of buying a single bundled stack from one AI lab. It sounds like an architecture debate. It is really a debate about who owns your negotiating position.
The context is a shift most companies are living through right now. AI is moving out of the prototype phase and into production, and production changes what matters.
Prototype Forgives Anything. Production Does Not.
In a prototype, you grab whatever model is easiest and wire it to whatever framework is nearest. Nothing is optimized because nothing needs to be. You are proving the idea works, and lock-in is a problem for later.
Later arrives when you ship. As Rauch puts it, when you are optimizing for production, you start looking at price and performance. Suddenly the model you grabbed for convenience is a line item you pay for at scale, and its price and speed on your actual workload are numbers that show up every month. The casual choice becomes an expensive commitment.
His argument is that this should work like the rest of software engineering, where your components come from different providers chosen for fit, not delivered as one take-it-all package. The agent, the layer that plans and orchestrates, should be separable from the model, the layer that generates. Decouple them and you can swap either one when a better or cheaper option arrives, across OpenAI, Anthropic, Gemini, and the open models climbing up from below.
Why the Labs Want the Opposite
The reason this is a fight is that the AI labs want the bundle. If the model and the agent framework come from the same vendor, you are not a customer choosing components. You are a tenant. Your orchestration logic, your workflows, and your model calls all live in one house, and moving out means rebuilding, which you will not do.
That is the vertical integration play, and it is rational for the labs. It is just not aligned with your interest. I wrote in a recent edition that your AI vendor can be switched off, and coupling your entire stack to one provider is how you make that switch impossibly expensive. Convenience today, captivity tomorrow, and the bill in between.
Vercel's answer is to build for the decoupled world. It made Eve, an agent framework that takes instructions in natural language, and Vercel Sandbox, which restricts what data an agent can reach so its actions stay auditable. Rauch describes Vercel as the AWS of this generation, fighting for open protocols where models and agents stay separate. Take the self-interest as given, agenda and all, and the underlying architecture argument still holds.
The Move for Everyone Building on AI
Most companies reading this are not building agent frameworks. They are deciding how to put AI into their operations, and the lesson lands the same. Two use cases are driving real production usage, coding agents that burn tokens and internal corporate agents that pull data together for sales, operations, and analysis. Both are becoming load-bearing, which is exactly when lock-in starts to hurt.
So design for portability from the start, even when it is slightly less convenient. Keep the model choice separate from the workflow logic, so switching models is a config change, not a rebuild. Route each task to the model that fits it, since most work does not need the frontier and paying frontier prices for commodity tasks is pure waste. Keep the orchestration, the part that encodes how your business actually runs, in your own hands rather than baked into a vendor's proprietary layer.
I made the case for vendor optionality a few editions ago, and the production shift only sharpens it. The model underneath you will get repriced, upgraded, or beaten by a cheaper option, probably more than once this year. If your stack assumes a single vendor forever, every one of those moves is a crisis. If your stack assumes swapping, every one is an opportunity to cut cost or gain capability.
The cheapest time to build that portability is before you need it. Retrofitting a decoupled architecture onto a stack already fused to one vendor is a migration project nobody funds until a price hike forces it, and by then you negotiate from weakness. A thin abstraction between your workflow and the model, added while the system is still small, costs almost nothing today and buys every future switch at the price of a config change.
The decoupling fight looks technical, and the labs would love for you to treat it as their problem to sort out. It is your negotiating position on the table. The company that keeps the model and the agent separate can always walk toward a better deal. The company that let one vendor own the whole stack can only hope the landlord stays generous. Build for the version where you get to leave.