How we Solved Agent Building — Andrew Qu, Vercel
Andrew Qu from Vercel started with an internal AI agent to solve a problem: data scientists became the bottleneck when other teams wanted answers about customers and products.
His first attempt was a large prompt with the database schema pasted in — it worked poorly. He then tried a chain of narrow agents for planning, execution, and reporting, and later a single agent managing its own state. The breakthrough came when he saw that Claude Code could answer the same questions without missing a beat, using simple tools like a file system — not advanced integrations. He rebuilt the agent with a sandbox and a semantic layer, and the result was twice as good. Vercel then released Eve, a framework using Next.js conventions for agents, and now has about twenty specialized agents in use.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
