What Is Vercel Eve — and What It Means for Building Agents

In June 2026, Vercel released Eve, an open-source framework for building, running, and scaling AI agents. Its pitch is deliberately familiar: “Next.js for agents.” Here's what it actually is, why long-running agents are the hard part it solves, and how we recommend using it.

An agent is just a directory of files

Eve's core idea is filesystem-first. You don't wire up a graph or subclass an orchestrator — you create a directory, and the files in it are the agent:

• • Instructions — the agent's role and rules, written in markdown (instructions.md).
• • Skills — reusable markdown playbooks that give the agent focused, on-demand guidance.
• • Tools — TypeScript files that automatically become callable functions.
• • Config — model choice and settings in agent.ts.

The slogan is “markdown for instructions and skills, TypeScript for tools.” You start with npx eve@latest init my-agent and build up incrementally by adding files. It's open-source under Apache-2.0 and published as the eve npm package.

The real unlock: durable, long-running agents

Most agent frameworks are great at a single request-response loop and fall apart the moment an agent needs to run for hours, wait on a human, or survive a deploy. Eve is built for exactly that. Sessions run on top of Vercel Workflow, which persists progress as an event log and deterministically replays it to reconstruct state. The practical result: a session can survive cold starts, redeploys, and indefinite pauses, then pick up exactly where it left off.

That changes what you can build. Human-in-the-loop approvals are first-class: any action can be set to require confirmation, and the agent pauses and waits without consuming compute until someone approves — then continues. A scheduled agent can run for days, escalate to a person, sit idle for a week, and resume cleanly.

Production is built in, not bolted on

Beyond durability, Eve ships the things you'd otherwise assemble yourself:

• • Sandboxed compute — agent-generated code is treated as untrusted; every agent gets its own sandbox (Vercel Sandbox in production; Docker, microsandbox, or just-bash locally).
• • Subagents — delegate specialized work to child agents.
• • Schedules — run agents on cron automatically.
• • Channels — deploy one codebase to Slack, Discord, Teams, and web chat.
• • Connections — managed auth for services like GitHub and Stripe.
• • Evals — test suites and scoring rubrics to measure agent quality.

What it means for the ecosystem

2026 has been the year agent frameworks split into camps. LangGraph remains the production standard for explicit, stateful Python workflows. CrewAI owns fast role-based multi-agent crews. The Microsoft Agent Framework unified AutoGen and Semantic Kernel for the enterprise. Mastra brought a TypeScript, web-native option.

Eve's contribution is to make the operational hard parts — durability, sandboxing, approvals, deployment — the default, while keeping the authoring model as simple as editing files. If you live in the TypeScript and Vercel world and you're building agents meant to run in production over time (not just demo in a notebook), it's the most direct path we've seen.

How we recommend using Eve

On AgentBlueprint, Eve is now one of our two primary recommended routes for building agents, alongside OpenClaw. Reach for Eve when an idea involves long-running work, scheduled runs, human approvals, or production deployment — which describes a large share of our blueprints, from outreach and ads agents to monitoring, reconciliation, and support automation. Many of those map almost one-to-one onto an Eve agent directory plus a schedule.

Every blueprint on the site lists its recommended frameworks and ships a copy-paste prompt, so you can scaffold with npx eve@latest init and drop the prompt straight into your agent IDE.