Vercel's Andrew Qu on why agents are a new kind of software

Andrew Qu has a job title that sounds made up until you hear him explain it: Chief of Software at Vercel, working alongside the CTO on whatever needs building next. Lately that has meant agents, and Qu has been in the thick of it, from writing an early MCP library to launching skills.sh, a discovery tool for agent skills, to leading eve, Vercel's in-house framework for building agents.

The path there was not a strategy memo. It started with v0, Vercel's vibe-coding product, and the team hit the same annoying problems over and over: swapping models, adding fallbacks, making long runs resumable without losing state. Those fixes turned into libraries, the libraries turned into shared primitives, and about a year ago Qu set himself a goal of putting an agent on every desk inside Vercel. A data agent came out of that effort, along with a batch of now-familiar patterns — filesystem agents, skills, compaction, subagents — that Qu says he wishes had simply come out of the box. Eve is Vercel's attempt to hand those patterns to everyone else instead of making each team reinvent them.

Qu's central claim is that agents are not web apps with extra steps. The infrastructure might look familiar, but the behavior is genuinely less predictable, and that unpredictability forces different engineering choices around context, tools, and long-running work. He is not selling full autonomy as the answer either. His rule of thumb: if a task is well-scoped and you already know what success looks like, let the loop run to completion; if the work is delicate, keep a human checking in along the way.

Skills, in Qu's telling, solve a subtler problem — model knowledge going stale. He points to Vercel Postgres, a product the company killed years ago in favor of a marketplace approach, which language models still cheerfully recommend because their training data never caught up. A skill file can correct that gap on the fly, which matters more the longer companies wait to audit old documentation.

The bigger structural shift Qu describes is the web quietly splitting in two. Vercel already detects when a request comes from a bot or agent and serves Markdown instead of HTML built for human eyes, because bot traffic is climbing while human traffic has flattened. Humans get the visual site; agents get something closer to raw, structured data. And Vercel's own ambitions point the same direction — the company now has agents embedded in its website, Slack, and dashboard, with Qu framing the long-term goal less as shipping agent products and more as making the entire platform behave like one.