Andrew Qu is Chief of Software at Vercel, working with the CTO on internal engineering, product experimentation and emerging technologies. He has built libraries for MCP, created skills.sh, and led development of eve, Vercel’s framework for building agents.
From web apps to agents
Vercel started out simplifying deployment for websites and web applications. In recent months the company has seen developers move from building pages toward building agents. While developing an early internal agent (v0, a vibe-coding product), the team encountered numerous minor but blocking issues that existing tools did not address: switching models or providers, adding fallbacks, and making runs resumable.
Those solutions were turned into reusable libraries to support v0 and to help customers build their own agents. Over time the collected primitives — skills, filesystem agents, subagents, compaction techniques and more — were assembled into a more cohesive framework called eve.
Why a dedicated agent framework became necessary
About a year ago Qu began working toward putting an agent on every desk inside Vercel. Building a successful data agent surfaced a set of best practices: filesystem agents, skills, compaction, and subagents — features Qu wished had been available out of the box. To spare other developers the same exploratory work, Vercel created eve as a prescriptive approach for building agents.
Agents as a distinct software category
Qu argues that agents are a new type of software. They are less predictable than web applications: interaction patterns and outputs are much more dynamic even if the underlying infrastructure looks similar. That demands different primitives for managing context, integrating tools, supporting resumability, and handling long-running work.
Suitable use cases for agents
Internally at Vercel, agents are used for repetitive, reasoning-heavy tasks such as first-pass legal contract redlining, marketing retrospectives, identifying people to contact, and writing queries against company data stores. Good candidates are repetitive tasks that still require interpretation and decision-making rather than fixed automation.
Autonomy vs human-in-the-loop
Qu does not see a future that is purely autonomous nor purely human-in-the-loop. The right approach is choosing a feedback cycle that fits the task. If a task is well-defined and the expected output is known, running an autonomous loop can be appropriate. For careful or surgical engineering work, periodic human checks are advised to ensure the model is being steered correctly.
Lessons from building eve
Vercel is still discovering what makes agents productive. They have been collecting primitives and bringing them together in eve, but more components will be added as best practices emerge. A year ago the team did not anticipate how important sandboxes, secure code execution, and support for long-running jobs would become.
When you deploy eve on Vercel, observability and evaluation features are available out of the box. Vercel intends to keep the developer experience easy while enabling integration with specialized partners rather than owning every lifecycle component themselves.
Why skills matter
Skills act as portable, on-demand knowledge that can correct model outputs when base models contain outdated information. Qu notes models sometimes still recommend Vercel Postgres, even though Vercel deprecated it years ago in favor of its marketplace. A skill can inform an agent that Vercel Postgres is deprecated and direct it toward the current approach. Qu recommends publishing skills for the latest product version and auditing existing content to update or annotate outdated material.
An agent-readable web
Vercel has reported rising bot traffic while human traffic is flat or falling, meaning agents and bots are hitting websites more often. The web’s future is to be accessible to bots and agents so they can learn about products and use them successfully.
Vercel already detects agent requests and serves Markdown directly instead of HTML designed for visual browsers, providing a machine-friendly representation. That points toward diverging experiences: humans receive the visual site, while agents receive a structured, machine-readable format.
What’s next: multiplayer development and platform-as-agent
Qu is focused on multiplayer agent development: sharing context among team members is a frequent challenge, and he’s exploring ways to let teammates contribute to and share that context. Whether agents become a distinct application category or a standard capability built into most software depends on the organization and product. For Vercel, the platform itself is becoming agent-like — there are agents on the website, in Slack, and in the dashboard that can act on users’ behalf, and the company aims to make the entire platform agent-friendly.
