Vercel Eve — Full Course: Build & Deploy AI Agents (End-to-End)

1. What Is Eve? The Next.js for Agents

Eve treats durable agents as ordinary files in a TypeScript project, following the same philosophy as Next.js: app router maps to agent folders. It's file-system-first and durable by default.

Agent folder structure:

• agent.ts + instructions.md — required
• tools/ — custom tool definitions
• skills/ — on-demand knowledge & procedures
• channels/ — multi-surface channel configs
• connections/ — MCP server integrations
• sandbox/ — isolated code execution
• sub-agents/ — delegated specialist agents
• schedules/ — cron-like automation triggers

2. Setup & Scaffolding

Prerequisites: Node.js 24+ (use NVM for version management).

Uses the Vercel AI Gateway for model connections. Set model credentials in your environment variables.

Two essential files:

• agent.ts — model configuration (which LLM, options, cost tracking)
• instructions.md — the agent's system prompt / permanent identity

3. Adding Tools

Create a tools/ folder inside the agent directory. Each tool is one TypeScript file.

Use defineTool from eve/tools with Zod for input validation. Every tool needs:

• description — what the tool does (for the LLM)
• input schema — Zod schema for parameter validation
• execute function — the actual logic

4. Agent Configuration

agent.ts defines the model, options, and cost tracking. Instructions form the agent's permanent identity, loaded every single turn.

Length guidance: 1–2 lines for simple agents, 10–30 lines for medium complexity, 50+ lines → split into skills.

5. Instructions vs Skills

This is a critical distinction in Eve's architecture:

• Instructions — loaded every turn. Permanent identity, tone, and rules.
• Skills — loaded on-demand when the model calls load_skill.

Skill formats:

• Markdown file (simplest)
• TypeScript (defineSkill)
• Split directory (for complex skills)

6. Channels: Multi-Surface Agents

One agent codebase, many surfaces. Every agent ships with a built-in HTTP channel. Add more via the Eve CLI:

The key insight: agent logic stays identical across all channels. The channel layer handles surface-specific I/O (message format, auth, webhooks) while the agent core remains unchanged.

7. Connections & MCP Servers

Connect external services via MCP (Model Context Protocol). The connections/ folder holds MCP server configurations.

• Agent can use tools from any connected MCP server
• Standard protocol — works with any MCP-compatible service
• Extends the agent's capabilities without writing custom tool code

8. Sandbox & Code Execution

Eve provides an isolated sandbox for agent code execution — the agent can write and run code safely.

• Sandbox configuration lives in the sandbox/ folder
• Supports file system access within sandbox boundaries
• Secure by default — prevents agents from accessing host resources

9. Durable State & Sessions

Durability is a first-class feature of Eve, not an add-on:

• Every turn and tool call is checkpointed
• Sessions survive crashes and redeploys
• Resume from any breaking point naturally
• State management is built into the framework

10. Sub-Agents & Delegation

Built-in agent tool: the model can delegate subtasks to copies of itself — same instructions, same tools, shared sandbox, but fresh history.

Delegation patterns:

• Parallel fan-out — multiple sub-agents tackle tasks simultaneously
• Specialist handoff — route to domain-specific sub-agents
• Research-then-act (ReAct) — sub-agent researches, parent acts on results

Task mode: sub-agent runs to completion, returns a structured result via output schema.

11. Schedules & Automation

Define scheduled tasks in the schedules/ folder for autonomous, recurring agent execution.

• Agent runs autonomously on schedule — no user trigger needed
• Cron-like scheduling for recurring tasks
• Each schedule is a separate configuration file
• Combines naturally with tools, skills, and sub-agents

12. Deploy to Production

Deploy via Eve CLI or the Vercel plugin:

Production features:

• Scales to thousands/millions of users
• Evals as deploy gate — test before shipping
• Smoke tests run automatically on deploy
• Durable sessions in production
• Auto-scaling and monitoring built in