With Vercel's eve, your AI agent is a directory

With eve, your agent is a directory

Vercel's eve makes a bet that sounds almost too simple: an AI agent should be a folder. Instead of wiring an agent together in code, eve puts its core capabilities in conventional files and directories, so an agent is something you can inspect, diff, and operate the way you already handle the rest of a project. It is a filesystem-first framework for durable agents, currently in beta.

The structure is the whole idea. An eve agent lives under an agent/ directory: instructions.md holds the always-on system prompt, tools/ holds typed functions the model can call, skills/ holds procedures loaded on demand, channels/ connects message surfaces like HTTP, Slack, or Discord, and schedules/ holds recurring cron jobs. Each capability has an obvious home, which means a new engineer can read an agent's behavior by reading its folder rather than tracing it through a framework.

A working agent in a few files

Getting started is one command: npx eve@latest init my-agent scaffolds the directory, installs dependencies, initializes Git, and starts an interactive terminal UI. A minimal agent is genuinely minimal. You write an instructions.md, drop a typed tool in tools/ defined with defineTool and a Zod input schema, and pick a model in agent.ts with defineAgent. That is a working agent, and you extend it with human-in-the-loop prompts, subagents, and schedules from there.

One detail signals who this is for. The eve npm package ships its full documentation inside node_modules/eve/docs, so a coding agent working in your repo can read eve's own docs locally. A framework that assumes another AI agent will be reading and editing its projects is designing for how this work increasingly gets done.

Where it fits, and the caveats

The caveats are the ones any beta carries, and eve names them. It is under the Vercel public beta terms, and the framework, APIs, documentation, and behavior may change before general availability, so it is a place to prototype and learn rather than to anchor a production system today. The typed-tool, Zod-schema approach will feel natural to a TypeScript team and foreign to anyone hoping to stay in Python.

If you have built an agent the usual way and ended up with a tangle that is hard to hand off, eve's premise is worth an afternoon: scaffold one, look at how little is hidden, and decide whether "the agent is a directory" makes your next one easier to operate. The idea is small, which is the point. Most of the friction in running agents is not the model, it is everything around it, and eve is an argument that the filesystem is a better place to put that everything than a framework's internals.

GitHub: https://github.com/vercel/eve

Curated by Agent Palisade — practical AI for small and mid-sized businesses.