Andrew Qu is Chief of Software in Vercel’s Office of the CTO and the creator of the installation CLI behind skills, its open agent-skills ecosystem. He builds software that makes organizational knowledge usable by AI agents, connecting internal experiments with developer tools for retrieving context, distributing expertise, and controlling how agents interact with external systems.
From collaboration software to agents
Qu’s startup experience began with Nooks, which he co-founded in 2020. Its early product helped people collaborate remotely through virtual rooms, video conversations, and shared documents. He subsequently worked as a software engineer at Meta from 2021 to 2023, explored and built with startups between 2023 and 2025, and joined Vercel in 2025. His career history spans founding a product, engineering inside a large technology company, and developing new infrastructure in Vercel’s Office of the CTO.
At Vercel, one consequential experiment began with a routine interruption: colleagues needed answers about customers and products, and analysts had to stop their work to write queries. Qu worked with the company’s data leadership on d0, an internal data-science agent that turns questions into SQL and answers. The goal was to give colleagues useful access to company data while freeing the data team to improve the underlying systems.
The architecture changed repeatedly. An initial large prompt gave way to a chain of specialized agents, then to one agent that could revisit its earlier work and recover from mistakes. The decisive simplification put the company’s semantic-layer documentation into a filesystem and let the model explore it with familiar shell commands. Rather than anticipating every possible question through custom retrieval tools, the agent could read definitions and relationships as it needed them. Qu’s filesystem-agent approach depended on well-documented data: direct access would accomplish little if measures, dimensions, and joins were unclear.
Making expertise available when it matters
Qu’s work develops several connected ideas:
Filesystem-based context retrieval. Large prompts can bury useful information and consume the model’s context window before the task begins. Qu and Malte Ubl introduced the open-source bash-tool, which lets AI SDK agents search files and retrieve selected results through shell-style commands. Its bash, file-reading, and file-writing tools work with in-memory and sandboxed environments. The mechanism keeps a larger body of information available without loading all of it into every model request.
Reusable agent skills. Successful work should give the next agent run a better starting point. For d0, recurring query patterns became skills containing useful company context and procedures. Qu also helped turn Shu Ding’s accumulated React expertise into something coding agents could readily consume. Repeating installation instructions for different agents led Qu to build the CLI that became npx skills; installation telemetry then supplied data for the skills.sh directory and leaderboard. His contribution addressed distribution: making expertise easy to install across agents, while preserving credit for the people who supplied it.
Versioned tool definitions. An agent’s behavior depends partly on tool names, descriptions, and argument schemas. Qu and Ubl developed mcp-to-ai-sdk to generate local AI SDK tool definitions from MCP servers. Teams can review, customize, and version those definitions instead of accepting upstream changes automatically. This limits unexpected changes to the agent’s available capabilities and instructions; responses from the remote server still require treatment as untrusted input.
Tools organized around user intentions. With Boris Besemer, Qu argued for MCP tools that complete meaningful tasks, rather than simply reproducing low-level API operations. Otherwise, each new conversation forces the model to rediscover the same sequence of calls. This complements his filesystem work: give the model room to investigate, and make external actions understandable in terms of the outcome the user wants.
More recently, Qu co-authored the introduction of the Vercel plugin for coding agents. The plugin uses signals such as file edits, terminal commands, and imports to select relevant platform knowledge, deduplicate it, and control how much enters the agent’s context. His work on company-specific agents also extends toward Eve, Vercel’s agent framework, which packages skills, tools, and communication channels through filesystem conventions. The ambition is to let other builders begin with the lessons learned from d0: accessible knowledge, reusable procedures, and infrastructure that supports the agent through a complete task.
Andrew Qu follows Vercel’s data agent from a pasted Snowflake schema through scoped agents, a sandboxed file system, reusable skills, and Eve. Each rewrite changes how the agent finds context, recovers from mistakes, and uses company knowledge.
Automate the complete question-to-answer workflow: interpreting data relationships, executing and revising SQL, and explaining the result.
The file-system redesign made the semantic layer searchable inside a sandbox, using familiar file operations and bash alongside a few Vercel-specific tools.
Eve packages skills, tools, and channels through file conventions; company knowledge still determines whether the resulting agent makes useful business decisions.
