Eve by Vercel: A Filesystem-First Framework for Durable AI Agents

🤔 The Problem: Framework Fatigue in AI Agent Land

You know the feeling. You want to build an AI agent — something that can use tools, remember context, call APIs, maybe run on a schedule. So you pick a framework. But most of them hit you with a wall of config: YAML files, state machines, orchestrator services, vector stores, message queues... Before you've written a single line of agent logic, you're debugging Docker networks.

That's where Eve by Vercel flips the script. Instead of abstracting the filesystem away, Eve puts it front and center. Your agent's brain lives in plain directories and markdown files. It sounds crazy simple — and that's the point.

📁 What Makes Eve Different?

Most agent frameworks are built around runtime state — a complex orchestrator that manages conversation history, tool registration, and execution flow. Eve says: the filesystem is your state. Every capability of your agent lives in a conventional file or directory:

my-agent/
└── agent/
    ├── agent.ts            # Model config (which LLM to use)
    ├── instructions.md     # The always-on system prompt
    ├── tools/              # Typed functions the model can call
    │   └── get_weather.ts
    ├── skills/             # Procedures loaded on demand
    │   └── plan_a_trip.md
    ├── channels/           # Message channels (HTTP, Slack, Discord)
    │   └── slack.ts
    └── schedules/          # Recurring cron jobs
        └── weekly_recap.ts

No YAML. No orchestration service. No hidden state. Want to know what your agent can do? Look at the files. Want to add a tool? Create a file. Want to change the system prompt? Edit a markdown file. It's refreshingly boring — and that's a compliment.

⚡ Key Features — The Short List

I've been playing with Eve for a few days, and here's what actually stood out (not just marketing fluff):

🧠 Filesystem-First Architecture

This is the big one. Because agents are just files, you can git-track everything, review PRs on agent behavior, and even fork agents like code. I accidentally broke my agent's instructions.md once and just git reverted it. Try that with a database-backed agent platform.

🔧 Typed Tools with Zod

Tools are TypeScript functions with Zod input schemas. This means your LLM gets perfect JSON schema for tool calls, and TypeScript catches your mistakes at compile time. No guessing what shape the model expects — it's right there in the type.

📡 Multi-Channel Out of the Box

Your agent can listen on HTTP, Slack, or Discord with zero infrastructure. I set up an HTTP channel in about 30 seconds — Eve exposes a REST endpoint that streams your agent's responses. The Slack integration took maybe 5 minutes.

⏰ Built-in Scheduling

Need your agent to run a weekly report? Create a schedules/weekly_recap.ts file with a cron expression. That's it. No separate job runner, no sidecar services.

📊 How Eve Stacks Up

I've tried most agent frameworks. Here's my honest comparison:

Bottom line: If you know TypeScript and want an agent you can actually understand, Eve is the clearest option I've found. It's not the most feature-packed (no built-in RAG, no vector store), but what it does — simple, inspectable, durable agents — it does beautifully.

🎯 Is Eve For You?

Yes, if:

• ✅ You write TypeScript and want a framework that respects your workflow
• ✅ You want agent config that's reviewable in PRs (no clicking around a UI)
• ✅ You need your agent to talk to Slack, Discord, or HTTP without extra infra
• ✅ You value simplicity over "one framework to rule them all" complexity

Maybe not, if:

• ❌ You need a visual agent builder or no-code interface
• ❌ You want built-in RAG, vector search, or knowledge bases
• ❌ Python is your primary language (Eve is TypeScript-first)

💭 Final Thoughts

Eve is one of those tools that feels obvious in hindsight. Of course agent configuration should live in the filesystem — we've been managing code this way for decades. Vercel's bet is that the best interface for AI agent development is the same interface developers already use every day: a file tree and a text editor.

After a week with it, I'm convinced. Eve's approach makes agents feel less like fragile magic and more like regular software — something you can understand, debug, and ship with confidence. If you've been frustrated by over-engineered agent frameworks, give Eve a try. You might be surprised how far simplicity can take you.