Andrew Qu

Andrew Qu is Vercel’s Chief of Software, working with its CTO across internal engineering, product experimentation and emerging technologies. He built an MCP library, created skills.sh and led development of eve, Vercel’s framework for building agents. In his July 2026 interview, he argued that agents require different software primitives because their interactions, interfaces and outputs are more dynamic than those of web applications.[1]

Qu traced eve to internal experimentation: roughly a year before that interview, he began working toward putting an agent on every desk at Vercel and built a data agent. Filesystem agents, skills, compaction and subagents emerged as practices he wished had been available from the start. His work informed the team’s decision to assemble a more prescriptive framework so other developers could avoid repeating that exploration.[1]

Qu’s earlier accounts of d0 and the skills CLI show the concrete design choices behind that framework work: simplifying how an agent explores data and making expertise easier to distribute.[4][5][1]

In his December 22, 2025 account, Qu described a team redesign of d0, an internal agent that answers Slack questions by translating natural language into SQL. Its original architecture used specialized schema tools, constrained prompts, validation logic and hand-coded retrieval. Maintaining that scaffolding required repeated patches and adjustments as models changed.[4]

The replacement let Claude Opus 4.5 explore a Cube semantic layer stored as YAML, Markdown and JSON files in Vercel Sandbox. Standard commands such as grep, cat, find and ls exposed dimension definitions, measure calculations and join relationships directly, reducing the need for custom tools that summarized them. The displayed implementation retained SQL execution alongside command execution.[4]

Qu reported that, across five representative queries, successful results rose from four to five, average execution time fell from 274.8 to 77.4 seconds, and approximate average token use fell from 102,000 to 61,000 per query. He attached a material condition to those results: the semantic layer was already well documented, consistently named and clearly defined. Raw file access, he warned, would not rescue undocumented joins or confusing legacy naming.[4]

Qu’s February 20, 2026 account credited colleague Shu Ding with writing down his React and browser expertise during a weekend the previous year. Qu considered an MCP server too heavy for this collection of Markdown documents; skills offered a quick way to deliver the knowledge on demand. After repeatedly writing installation instructions for coding agents with different directories, Qu built the CLI that became npx skills.[5]

The team added installation telemetry that supplied the skills.sh leaderboard’s data. Qu credited Vercel CTO Malte Ubl with framing the project as a package manager for agent context. His broader lesson was that relevant context matters more than its format: code examples, inline documentation and MCP hints can also provide the patterns an agent needs.[5]

Qu advocates matching an agent’s feedback cycle to its task. In July 2026, he said a well-defined task with a clear final output may justify letting a loop run to completion, while careful engineering work calls for checking in and steering the model. His position leaves room for both autonomy and human involvement rather than prescribing one mode for every job.[1]

He described skills as “portable, on-demand knowledge” that can correct stale model recommendations. His example was models continuing to recommend deprecated Vercel Postgres: a skill could redirect them toward the marketplace approach. He also urged companies to update outdated content or add clear notes, making skills one part of maintaining current product knowledge.[1]

Qu also favors making websites readable by agents. In the same interview, he said Vercel detected agent requests and served Markdown directly rather than requiring agents to process a visual site’s HTML. He identified multiplayer agent development as a priority, particularly helping teammates share and contribute the context that makes their agents effective.[1]

Before the Vercel work described above, Qu’s personal career record lists co-founding Nooks in 2020–2021, working as a software engineer at Meta in 2021–2023, and building and exploring at various startups in 2023–2025. He dates his Chief of Software role in Vercel’s Office of CTO to 2025.[3]