Vercel Launches Eve Framework for Defining AI Agents as Directories

Vercel has introduced Eve, a new framework for building AI agents that treats an agent as a simple directory structure. Similar to how Next.js simplifies web apps, Eve uses Markdown for instructions and skills, and TypeScript for tools, aiming to make agent development durable and straightforward.

Directory-Based Architecture

An Eve agent is defined by a folder containing specific files:

• instructions.md: Describes the agent's role and identity.
• agent.ts: Configures the runtime and selects the model via AI Gateway.
• tools/: Contains TypeScript files where the filename becomes the tool name, requiring no registration.
• skills/: Holds Markdown playbooks loaded only when relevant to the task.

Integrated Infrastructure

Eve leverages existing Vercel primitives to handle complex infrastructure needs out of the box:

• Vercel Sandbox: Provides isolated execution environments for each agent.
• Vercel Workflows: Ensures durable execution, allowing agents to park and resume across crashes or restarts.
• Vercel Connect: Manages authentication for services like GitHub, Stripe, and Linear, enabling tools to call APIs without manual token management.
• Chat SDK: Supports multi-channel delivery to Slack, Discord, Microsoft Teams, WhatsApp, and web chat.

Production Features

The framework includes enterprise-grade capabilities such as human-in-the-loop approval gates, subagents for specialized tasks, and scheduled runs for automated jobs. Developers can deploy their first agent to Vercel in less than 60 seconds using the npx eve@latest init command.