Andrew Qu traces Vercel's effort to reduce the data team's repeated work answering internal questions. The first prototype placed a Snowflake schema and a question into one large prompt, then relied on a human to run the generated SQL.
The team next divided the workflow among planning, query, execution, and reporting agents with narrow tools. That architecture completed more of the process end to end, but summaries passed between agents discarded useful context and made unfamiliar requests difficult to handle.
A single agent with broad context and the ability to manage its own state performed better because it could revisit earlier work, recover from execution errors, and decide when to plan, explore, execute, or report. Moving the semantic layer into a sandboxed file system also let the agent use familiar operations such as reading files, searching, writing, and running commands.
Repeated internal requests were then distilled into reusable skills, giving later runs established knowledge instead of forcing every agent to start from an empty context. The talk presents these skills as a practical way to capture recurring query patterns while leaving the agent room to explore.
The broader lesson is that business-specific knowledge can matter more than an off-the-shelf vertical agent. Vercel reports that tailored agents now support data, marketing, sales, finance, and legal work, freeing specialists to improve systems rather than repeatedly handle routine requests.
Watch on YouTube
