Building Enterprise Agents with File-System Conventions

The File-System Unlock

Andrew Qu discovered that moving away from prescriptive agent chains toward a file-system-based sandbox significantly improved performance. By providing the model with a minimal toolset—list, read, and bash—the agent could explore and execute tasks with emergent behavior similar to Claude Code. This architecture allowed the agent to manage its own state, reflect on errors, and retry steps without requiring hard-coded branching logic.

Scaling Knowledge with Skills

To handle recurring data requests, the team implemented a distillation process that converts common queries into reusable "skills." These skills act as pre-established context, allowing new agent runs to start with a library of roughly 100 domain-specific capabilities rather than from scratch. This approach ensures that agents leverage company-specific knowledge, which Qu argues is superior to general-purpose vertical agents for internal enterprise workflows.

Eve Framework

To standardize agent development, Vercel released Eve, a framework that applies Next.js-style file-system conventions to agentic infrastructure. Developers define agents by organizing files into specific directories for skills, tools, and channels. The framework handles runtime durability, isolated execution environments, and observability out of the box, allowing teams to deploy custom agents that integrate directly with existing company data layers.

Before / After

• Initial Prototype (Mega Prompt): Low confidence, required manual SQL execution.
• Version 2 (Chained Agents): Improved but hit architectural walls due to limited context passing between agents.
• Version 3 (File-System Agent): Eval scores doubled compared to the chained architecture, enabling reliable end-to-end execution.