Vercel recently introduced Eve, an open-source framework (licensed under Apache-2.0) designed to standardize how we build, deploy, and scale AI agents. Pitching it as “Next.js for Agents,” Vercel is replacing the tedious infrastructure plumbing of agent design with an elegant, convention-over-configuration filesystem layout.
Instead of writing complex graph orchestration code, an Eve agent is defined simply by a directory of structured files. When compiled, Eve wires up durable execution workflows, registers tools dynamically, isolates execution sandboxes, and exposes communication channels out-of-the-box.
🛠️ The Architecture: An Agent is a Directory
Just as Next.js maps files in a pages/ or app/ directory to web routes, Eve maps files inside an agent/ directory directly to agentic capabilities.
my-agent/
├── agent/
│ ├── instructions.md # The always-on system prompt
│ ├── agent.ts # Runtime configuration (Model & AI Gateway)
│ ├── tools/ # Directory for TypeScript tools
│ │ └── get_weather.ts
│ ├── skills/ # On-demand Markdown procedures
│ │ └── research.md
│ ├── channels/ # Surface integrations (Slack, Discord, Web)
│ │ └── slack.ts
│ └── subagents/ # Specialized child agents
└── package.json
Step-by-Step Implementation Guide
Setting up a production-ready agent with Eve requires minimal boilerplate. Here is how to initialize and structure your first agent project.
Scaffold a fresh project using the Eve CLI. This initializes dependencies, generates the folder structure, and installs the eve npm package.
npx eve@latest init my-agent
cd my-agent
Define your base model and runtime behaviors. Eve routes these requests through Vercel’s AI Gateway, handling fallback strategies and tracking token limits automatically.
// agent/agent.ts
import { defineAgent } from "eve";
export default defineAgent({
model: "openai/gpt-5.4-mini", // Resolves via Vercel AI Gateway
});
Create the permanent personality and boundary parameters for the agent. This acts as the global system instructions.
# Role
You are an autonomous operations assistant tasked with monitoring system health and alerts.
# Constraints
- Never share infrastructure credentials.
- Always fetch the latest weather or system status before reporting metrics.
Drop a typed TypeScript file inside agent/tools/. The filename automatically becomes the tool name exposed to the LLM—no manual registration or manifest mapping required.
// agent/tools/get_weather.ts
import { defineTool } from "eve/tools";
import { z } from "zod";
export default defineTool({
description: "Get the current weather for a specific city",
inputSchema: z.object({
cityName: z.string(),
}),
async execute(input) {
const res = await fetch(`${process.env.WEATHER_API_URL}/current?city=${input.cityName}`);
const data = await res.json();
return data.current_condition[0];
},
});
Core Production Features Reference
Eve distinguishes itself from existing frameworks (like LangGraph or CrewAI) by bundling enterprise infrastructure layers directly into the compilation step:
| Feature | How It Works | Behind the Scenes |
| Durable Execution | Conversations run as checkpointed state loops. If a server crashes, restarts, or undergoes a redeployment, the session resumes exactly where it left off. | Built on top of Vercel Workflows |
| Sandboxed Compute | Code generated or executed by the agent runs in isolated environments to avoid malicious runtime vulnerabilities. | Runs locally via Docker; maps to Vercel Sandbox in production |
| Human-in-the-Loop | Specific destructive actions or tools can be configured to halt execution until a human operator grants approval via a dashboard hook. | Pauses compute execution state indefinitely without timeout costs |
| Multi-Surface Channels | The exact same agent logic can respond simultaneously to web APIs, chat widgets, or corporate messaging apps. | Leverages Vercel Connect Chat SDK (Slack, Discord, Teams, GitHub) |
Advanced Concepts: Skills and Subagents
As agents grow, passing thousands of lines of prompts into every turn degrades accuracy and increases token overhead. Eve introduces two structural remedies for this:
1. Reusable Skills (agent/skills/)
Skills are on-demand Markdown playbooks. Instead of clogging the core system prompt, these files are loaded conditionally by the LLM only when relevant to the current user intent.
<!-- agent/skills/research.md -->
---
description: Run deep research on unfamiliar or changing technical topics
---
When the user's task is novel or ambiguous, follow this playbook:
1. Gather initial evidence using available search tools.
2. Cross-reference facts across multiple independent fetches before drafting an answer.
Delegated Subagents (agent/subagents/)
For distinct, heavy-duty processing separations (e.g., separating a broad user triage agent from a dedicated data compliance auditor), developers can drop standalone agent directories inside the subagents/ folder. Parent agents can hand off execution chains natively, merging the subagent’s payload back upon completion.
Local Development & Observability
To run your agent locally with an interactive, streaming terminal workspace UI:
eve dev
# To run your continuous integration (CI) evaluation suites against custom mock expectations:
eve eval
When you are ready to launch, running vercel deploy builds and ships the filesystem manifest straight into production, automatically provisioning your environment variables, background cron schedules, and Agent Runs observability dashboard tracks.