The Evolution of Agent Architecture

Andrew Qu’s team at Vercel initially struggled to build an effective data science agent (D0). Their first attempt—a single "mega-prompt"—provided a baseline but lacked reliability. They then moved to a chain-of-thought architecture where specialized agents handled planning, execution, and reporting. While this improved performance, it hit a ceiling because each agent only received a summary of the previous step, losing critical context.

The breakthrough came from observing the behavior of Claude Code. Instead of prescriptive, rigid tool chains, the team realized that providing a simple file system (list, read, write, bash) allowed the model to leverage its native training to explore and solve problems autonomously. By moving to a sandbox environment where the agent could interact with a full semantic layer via standard file operations, their evaluation scores doubled.

Scaling Knowledge with Skills

To move beyond generic capabilities, the team implemented a "skills" system. They identified that many internal data queries were repetitive in structure. They now run a recurring job that distills recent successful queries into roughly 100 reusable "skills." When a new agent run begins, it is injected with these skills, providing it with pre-established context and domain-specific knowledge rather than starting from scratch. This approach allows the agent to perform complex aggregations and lookups without needing to re-derive the logic every time.

Introducing Eve: The 'Next.js for Agents'

To help others avoid the trial-and-error process, Vercel released Eve, an open-source framework designed to standardize agent development. Eve applies the "file-system-as-framework" philosophy to agents, similar to how Next.js uses file conventions to define infrastructure.

Key components of the Eve framework include:

• Standardized Structure: Developers organize agents using skills, tools, and channels folders.
• Durability & Execution: Built-in support for Vercel Workflows for state management and sandboxed environments for secure code execution.
• Observability: Out-of-the-box tracking of agent runs, tool calls, and cost metrics.

Qu argues that the most effective agents are not general-purpose models, but narrow, company-specific agents stuffed with proprietary knowledge. Vercel currently uses roughly 20 such agents across departments, ranging from marketing retrospectives to legal contract redlining, which has significantly reduced the operational burden on their internal teams.