Build an Agent with Eve — The Open Agent Framework

1. Scaffolding an Eve Agent

Andrew and Casey, tech leads on Eve at Vercel, kick off this live Ship 26 NYC workshop. Eve was released just a few weeks prior at Ship London, and this session walks the audience through building a fully functional agent from scratch.

Getting Started with npx init

• A single npx init command scaffolds the most basic Eve application — minimal boilerplate, instant productivity
• Eve uses a declarative file-system-based approach to agents — no config DSLs, no sprawling setup

Running in Dev Mode

• eve dev walks through setup including model provider authentication
• Choose between AI Gateway (Vercel's managed gateway product) or bring your own keys/providers
• Out of the box you get a ChatGPT-in-a-box with web search and all harness features that "just work" ▶ 1:36

2. Channels — From Terminal to Slack

Channels are the topmost layer of how users interact with an Eve agent. They abstract away the presentation so the agent logic stays clean and channel-agnostic.

Channel Architecture

• A default HTTP channel runs behind the scenes for local dev ▶ 3:30
• Connecting to Slack via Vercel Connect takes seconds — deploy and the agent appears in your workspace
• The audience tags the bot live during the workshop — instant interaction in real-time ▶ 4:40

3. Debugging in the Vercel Dashboard

Eve provides full observability out of the box for every agent interaction, directly in the Vercel dashboard.

• Every message and run is viewable in the dashboard — see exactly what happened and what failed ▶ 6:20
• Debug tool call failures in real time — no need for external logging infrastructure
• Full trace of the agent's reasoning, tool invocations, and responses available per-interaction

4. Connections — OpenAPI Data in One Prompt

Connections are a first-class primitive in Eve for data integration. They let agents reach into external data sources with minimal configuration.

Connection Types

• OpenAPI specs — point at a spec, Eve generates the tools automatically ▶ 7:50
• Postgres databases — direct SQL access for data queries
• Any data source can be wrapped as a connection

Coding Agent Integration

• Used a coding agent (open code) to add an OpenAPI connection live during the demo ▶ 8:30
• Eve ships special documentation inside node_modules so coding agents can understand Eve concepts and scaffold correctly
• The agent immediately uses the new connection for weather queries

5. Tools with Approvals — The MongoDB Example

Tools in Eve are a one-to-one mapping for custom functionality. Drop them into a tools/ folder and they're immediately available — fully declarative.

MongoDB Query & Mutate Tools

• Separate query and mutate tools for read vs write operations ▶ 12:00
• Import any Node.js library you want inside tool calls — full ecosystem access
• Sandbox interaction available for safe execution

Human-in-the-Loop Approvals

• Approvals flow through the same UI per channel — native Slack buttons, terminal prompts ▶ 13:50
• Built-in question tool renders differently per channel (terminal vs Slack native UI)
• A WebSocket channel was built live by a coding agent during the demo — Eve has WebSocket as a channel primitive ▶ 16:00
• Vercel Connect team ships integrations (Linear, Slack, GitHub) separately from Eve core

6. Evals — Deterministic Checks & LLM as Judge

Eve includes a first-class evaluation system — critical for iterating on agents and understanding the impact of changes.

Two Types of Evals

• Deterministic checks — simple assertions like "does the response include the word NYC" ▶ 18:20
• LLM as judge — semantic evaluation of meaning and quality ▶ 19:00

• Can send eval results to providers like Braintrust for tracking over time
• The API mirrors the CLI testing interface — automate what you'd do manually in CI/CD

7. Frontend Frameworks & Generative UI

Eve is deeply integrated with frontend frameworks, with Next.js as a first-class citizen.

React & Next.js Integration

• useEveAgent hook for seamless frontend interaction ▶ 21:20
• Framework teams (SvelteKit, Nuxt, Next) all actively supporting Eve

Generative UI Demo

• World Cup demo: beautifully rendered generative UI components streamed directly from Eve ▶ 23:20
• Rich, interactive components generated on-the-fly based on agent responses

Eve vs AI SDK

• AI SDK v7 features are available same-day in Eve ▶ 26:00
• Eve handles durable execution, session management, and multi-channel delivery on top of AI SDK

8. Architecture — Stateless Harness on Workflow SDK

Under the hood, Eve's harness is fully serverless and stateless, using Vercel's Workflow SDK for durability.

Serverless Durability

• Workflow SDK provides retry semantics and stateless function dispatch ▶ 28:40
• Long-running turns can take hours — the system dispatches to a new serverless function at each tool call boundary ▶ 30:00
• Bypasses the 15-minute serverless function timeout entirely

Harness Features

• Default tools: read file, write file, web fetch, web search, glob, grep — all replaceable ▶ 32:00
• Hooks system: plug into harness events (compaction, pre/post tool calls) for custom behavior
• Every Eve agent has a sandbox — Docker locally, Vercel Sandbox in cloud ▶ 34:00
• Self-hosting story available in docs (different from managed Vercel deployment)
• Future: possibly run Eve on top of other harnesses (open code, Pi)

9. Eve vs The Field — Extensibility & Positioning

How does Eve fit alongside other agent products, including Vercel's own Tag?

Eve vs Tag

• Tag = pre-built Slack agent, single binary, self-evolution and memory
• Eve = code-first framework, build for your specific business ▶ 36:00

Progressive Complexity

• Multi-tenant APIs: per-user skills, memory stored in DB, scales to hundreds of thousands of users
• Every markdown file can be a TypeScript file — eject to custom logic at any point ▶ 36:50
• Skills in plain English: frontier models follow well-written instructions reliably
• Output schemas for deterministic behavior from non-deterministic LLMs
• Sub-agents for structured, composable output

10. Pricing, Browser Use & What's Next

Practical questions about costs, browser automation, and the Eve roadmap.

Pricing Model

• Deploying an Eve agent costs nothing (serverless) — pay only for usage ▶ 38:10
• Main costs: function CPU, function duration, sandbox usage
• Tokens account for ~90% of spend — the LLM calls dominate

Browser & Computer Use

• No built-in browser/computer use — plug into your favorite provider via tool calls ▶ 39:00
• Vercel Sandbox could support browser use in the future
• Built-in capabilities limited to web fetch and web search