March 30, 2026 · 4 min read

From zero to agent: prototyping with v0 and the AI SDK

At Vercel's Zero to Agent sessions I set myself one constraint: ship a working, type-safe AI agent without hand-writing the interface. The experiment was v0 as the prototyping environment, the Vercel AI SDK as the agent runtime, and MCP for tools. Here's what held up and what didn't.

The workflow

The loop that worked looked like this:

1. Describe the interface to v0, not the implementation. "A panel showing the agent's tool calls as they stream, with a retry affordance per step" gets a better result than pasting component code. v0 is strongest when you treat it as a design partner with opinions about React, not as autocomplete.
2. Own the agent loop in code. The AI SDK's tool-calling primitives stay in hand-written TypeScript — typed tool schemas, explicit control over what the model may call. Generated UI, hand-written behavior. That boundary is the whole trick.
3. MCP for the outside world. Wiring tools through Model Context Protocol servers meant the agent's capabilities were declared, inspectable and swappable — closer to configuration than code.

What surprised me

The generated TypeScript was more readable than I expected — enough that reviewing it felt like reviewing a colleague's PR rather than deciphering output. Accessibility held up too, but only when I asked for it explicitly. The lesson generalizes: generative tools meet the bar you set in the prompt, and silence is a decision.

What didn't hold up: state that lives across many components. Past a certain complexity, v0 iterations started fighting each other, and moving the state design back into my editor was faster than prompting around it.

Why this changed my default

Before, prototype quality and prototype speed were a trade-off. Now my MVP loop is: generate the interface, write the logic, and spend the recovered hours on the part that differentiates the product — for me, that's usually the domain logic where regulation meets code.

The workflow from that weekend is now how I start every new build, including parts of this site.