What Is Eve? A Quick Tour of Its Five Core Concepts

What Is Eve?

Eve is a file‑system‑first agent development framework. Like Next.js maps pages/ or app/ to routes, Eve maps an agent/ directory to an entire agent.

my-agent/
├── package.json
└── agent/
    ├── agent.ts          ← model and runtime
    ├── instructions.md   ← permanent identity
    ├── tools/            ← typed actions
    ├── skills/           ← on‑demand knowledge
    ├── channels/         ← Slack, Discord, Web, etc.
    ├── subagents/        ← delegated tasks
    └── connections/      ← external services

Five Core Concepts

1. Instructions

The agent/instructions.md file is injected at every turn as the agent’s permanent identity, suitable for role definition, behavior rules, and constraints.

2. Tools

Files under agent/tools/ define typed actions via defineTool. Input validation is expressed with Zod schemas, and the tool runs with full access to process.env.

// agent/tools/get_weather.ts
import { defineTool } from "eve/tools";
import { z } from "zod";

export default defineTool({
  description: "Get the current weather for a city.",
  inputSchema: z.object({ city: z.string().min(1) }),
  async execute({ city }) {
    return { city, condition: "Sunny", temperatureF: 72 };
  },
});

3. Skills

Markdown files in agent/skills/ are loaded on demand via the built‑in load_skill tool, preventing the prompt from being overloaded with static knowledge.

4. Channels

Eve ships with built‑in channel adapters for HTTP API, Slack, Discord, Web chat, Telegram, GitHub, Linear, and more. The same agent code runs unchanged across all channels.

5. Subagents

Complex sub‑tasks can be delegated to child agents defined under agent/subagents/. Each subagent has its own description, tools, and skills, and the parent decides when to invoke it.

Durable Sessions

Eve’s default “Durable by default” model adds a persistent session layer: streaming output, tool/subagent calls, human‑in‑the‑loop approvals, checkpoint‑based crash recovery, and cross‑turn state persistence.

Example: a refund tool marked with needsApproval: always() pauses the session; after a human confirms, the session resumes exactly where it left off.

How Eve Differs from Existing Solutions

Compared with LangChain/Graph (chain/graph abstraction, Python‑centric, manual persistence) and CrewAI/AutoGen (multi‑agent dialogue, manual persistence), Eve offers:

File‑system‑driven architecture (no YAML/JSON config).

Low learning curve for TypeScript/Next.js teams.

Built‑in multi‑channel support.

Out‑of‑the‑box durable sessions and human‑in‑the‑loop.

One‑click Vercel deployment (also supports self‑hosting).

Technology Stack Overview

Eve sits on Vercel’s AI infrastructure: AI Gateway (model calls, streaming), Vercel Workflow (persistent execution, checkpoints), Vercel Sandbox (isolated VM), Vercel Connect (OAuth/MCP), and Chat SDK (frontend UI).

Who Should Pay Attention

TypeScript/Next.js teams – zero switch cost.

Teams needing multi‑channel agents – same code for Slack, Discord, web, etc.

Production‑grade agent developers – durability, HITL, sandbox execution.

Vercel users – integrated deployment and monitoring.

Engineers who dislike YAML config – file‑system is the config.

When to Wait

Pure Python stacks – LangChain remains a better fit.

Needing a mature stable release – Eve is still in beta.

Complex graph orchestration – LangGraph may be more suitable.

Reluctant to bind to Vercel – self‑hosting is possible but less seamless.

Conclusion

Eve positions itself as “directory‑as‑Agent”: a file‑system‑first, TypeScript‑native, Vercel‑integrated framework that aims to be the “standard answer” for agent development, similar to how Next.js standardized web development.