Vercel's Andrew Qu: File Systems, Not Clever Tools, Unlocked Internal Agents — 20 Now in Production

AI Engineer · youtube · 2026-09-15

Vercel Chief of Software Andrew Qu shared the full iteration behind the company's internal data science agent, which became the open-source Eve framework:

Origin: interviewing marketing, sales, finance, and legal about their biggest pain point; the answer was the data team — every question meant a data scientist dropping everything to write a query, making them a company-wide bottleneck.
Three architectures: v1, one mega prompt with a pasted Snowflake schema and hand-run SQL; v2, a chain of scoped agents for planning/execution/reporting; v3, a single agent managing its own state. The first rollout was judged "awful" despite clearing 30% of evals.
The unlock: watching Claude Code answer the same questions flawlessly with a plain file system — list, read, bash — tools models already handle well. Rebuilding in a sandbox with the semantic layer dumped in doubled eval scores.
Skill distillation: a recurring job condenses common queries into 100 skills, so runs start with accumulated context. That became Eve, bringing Next.js file-system conventions to agents, released two weeks before the talk.
Takeaway: narrow agents stuffed with company-specific knowledge beat general ones; Vercel now runs about 20 internally.

Original post →

Vercel's Andrew Qu: File Systems, Not Clever Tools, Unlocked Internal Agents — 20 Now in Production

More from coding & agent