Summary

2026-07-04 · source · source ↗ · raw capture · wiki page

Malte Ubl, CTO of Vercel, interviewed at The Pragmatic Summit about building Vercel's internal data agent Dzero and the V0 product agent. The central thesis: "building agents is actually extremely easy and you don't need to buy them."

Dzero (Internal Data Agent)

A text-to-SQL engine in Slack. Started as many tools in a loop, rebuilt as a two-tool agent (bash + Execute SQL) — about 50 lines of code. Reads a YAML file describing every Snowflake column in plain English, then uses grep/tail to explore semantics and writes SQL.

Key architectural insight: As models get smarter, simpler agents work better. Complex tool loops become unnecessary. The YAML semantic layer is the durable artifact — it describes columns in plain English that the agent can grep through.

V0 Evolution

Evolved through four user bases driven by model capability leaps: 1. Frontend engineers (early models, Tailwind CSS workaround) 2. Backend engineers (could fix V0's mistakes) 3. Full-stack builder for non-engineers (Sonnet 3.5 pivot) 4. Tech-adjacent roles (PMs, designers, business people) building internal tools

The "Tailwind moment": Early models couldn't handle separate CSS files. Instructing them to use Tailwind (inline styles) succeeded because models could reason inline. This made V0 viable in 2023.

Framing as coding tasks: Models excel at coding because of training data. Non-coding problems (text-to-SQL) perform better when structured as coding tasks — "you get disproportional good results."

Organizational Practices

Engineering Career Impact

Automation at Scale

Key Metaphors

Full Transcript

The complete 34-minute transcript is available at the gist URL above. The interview was conducted by Gergely Orosz (The Pragmatic Engineer) at The Pragmatic Summit, February 11, 2026.