Vercel shipped something this week that deserves attention from anyone building AI agents: eve, an open-source agent framework built around a single, simple idea. An agent is a directory of files. Add a file, add a capability. No boilerplate registration. No infrastructure glue.
Released under Apache 2.0 on June 17, eve already has over 1,600 GitHub stars and comes from the team behind Next.js (github.com/vercel/eve).
What "Filesystem-First" Actually Means
In eve, every part of your agent has a conventional location:
my-agent/
└── agent/
├── instructions.md # System prompt, written in Markdown
├── agent.ts # Model and runtime config
├── tools/ # Typed functions the model can call
├── skills/ # Larger procedures loaded on demand
├── channels/ # Slack, Discord, HTTP, cron
└── schedules/ # Recurring cron jobs
Add a TypeScript file to tools/, and the filename becomes the tool name. No registry. No wiring. The framework discovers the directory structure at build time and compiles it into a running agent.
Most agent frameworks require registering tools in a central config, keeping registries in sync, and managing the mapping between files and what the agent can call. Eve eliminates that class of problem entirely.
Production Comes Built In
The more significant decision is bundling production infrastructure into the framework itself.
Every agent gets durable execution (sessions checkpoint every step, survive crashes and deploys), sandboxed compute (isolated VMs for code execution), human-in-the-loop approvals, subagents, and built-in evaluations — without the developer assembling these pieces separately.
Vercel describes eve as "Next.js for agents," and the comparison is apt. Next.js succeeded because it made the right production decisions by default — routing, rendering, optimization — so developers could focus on building. Eve does the same for agents: durability, sandboxing, and multi-channel delivery are framework concerns, not per-project assembly tasks.
Production Numbers, Not Demos
Vercel runs over 100 agents internally on eve, including a data-analysis agent handling tens of thousands of Slack queries per month and a support agent that autonomously resolves 92% of tickets. Those are production numbers.
The framework works with any model via Vercel AI Gateway (with provider fallbacks), any MCP server, and deploys with vercel deploy as an ordinary Vercel project. Support for other platforms is planned.
The agent framework landscape has filled up quickly — LangGraph, Mastra, CrewAI, Inngest AgentKit all compete for developers. Eve's differentiator is the filesystem-first design paired with built-in production durability. If you already deploy on Vercel, the integration is seamless.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.