How We Solved Agent Building — Andrew Qu, Vercel

When Andrew Qu handed Vercel's internal data science agent to its first trusted users, the verdict came back that it was awful. It had been clearing thirty percent of his evals and he thought the team was cooking. Qu is Chief of Software at Vercel, and he began a year earlier by asking marketing, sales, finance, and legal what they hated most about their jobs. The best answer came from the data team, lean and growing slower than the company around it. Whenever anyone had a question about a customer or a product, a data scientist dropped everything to write the query and report back. Qu walks the rewrites that followed: one mega prompt with a Snowflake schema pasted in and the SQL run by hand, a chain of narrowly scoped agents for planning, execution, and reporting, then a single agent managing its own state. The unlock came from watching Claude Code answer the same questions without missing a beat. The advantage was not a clever tool set but a plain file system with list, read, and bash, tools the models already handle well. He rebuilt the agent in a sandbox with the semantic layer dumped in, and the eval score doubled. A recurring job now distills common queries into roughly a hundred skills, so a run starts with accumulated context instead of nothing. That became Eve, the framework Vercel released two weeks before this talk, bringing the file system conventions of Next.js to agents. Qu closes by arguing for narrow agents stuffed with company specific knowledge over general ones, counting about twenty now in use across Vercel. Speaker info: - https://x.com/andrewqu - https://linkedin.com/in/andrew-qu - https://andrewqu.com Timestamps: 0:00 - Vercel and the AI SDK 1:35 - An agent on every desk 2:45 - Why the data team became the bottleneck 3:39 - Version one, one mega prompt 4:51 - Version two, a chain of scoped agents 5:58 - Version three, one agent managing its own state 7:17 - A rough first rollout, then the file system unlock 10:14 - Distilling recurring queries into skills 12:04 - Building the Next.js for agents 14:18 - Early users, results, and where to start