Summary
Cole Medin examines Eve, a new open-source AI agent framework released by Vercel that introduces a file-system-first architecture for building production-grade agents. Unlike personal agent setups, Eve is designed to scale to thousands of users while retaining the simplicity of a folder-based project structure — where every component (instructions, skills, MCP server connections, sub-agents, schedules, and channels like Slack or Discord) lives as a markdown or TypeScript file inside a single directory.
The framework’s standout feature is a compilation step that automatically traverses the folder, wires up all components into a unified manifest, and handles connections without requiring manual imports — analogous to how Claude Code processes skill files via slash commands. On the production side, Eve leverages Vercel infrastructure to provide durable sessions (turn-by-turn checkpointing that survives crashes and redeploys), isolated code execution sandboxes, and human-in-the-loop controls.
Medin demonstrates building a full Eve analyst agent using the official Vercel plugin for Claude Code, which ships pre-built skills and an MCP server to guide both development and deployment. The video also shows the agent running locally and deployed to Slack, connected to Claude Sonnet via the Anthropic API. For teams currently hand-rolling their own internal agent folder conventions, Eve offers a standardized, open-source alternative with production infrastructure handled out of the box.
📺 Source: Cole Medin · Published July 16, 2026
🏷️ Format: Deep Dive
