Vercel unveiled Eve on June 17 at its Ship conference in London, positioning it as the Next.js for AI agents. Eve is an open-source, Apache 2.0-licensed framework where every agent is a directory of files—a filesystem-first model that bundles production infrastructure directly into the framework rather than requiring teams to hand-roll it. Agents are written in TypeScript and Markdown, with a minimal setup of just two files (model definition and instructions).
Eve ships with six production capabilities built-in: durable execution with checkpointed workflows (via Vercel Workflow SDK), sandboxed compute on isolated VMs, human-in-the-loop approval gates, subagents for delegation, integrated evals for testing, and OpenTelemetry observability. The framework handles slack/queue-like persistence, so agents can pause between messages, survive crashes, and resume where they left off without losing context. Agents deploy to Vercel with vercel deploy, the same command as web apps.
Vercel says it already runs more than 100 internal agents on Eve (including d0 handling 30k questions monthly and Vertex resolving 92% of tickets autonomously). The framework supports any LLM provider via Vercel's AI Gateway, with model fallback built-in. For architects shipping agentic systems, Eve eliminates infrastructure boilerplate at deployment-time; the tradeoff is that frictionless hosting defaults to Vercel's platform (AI Gateway, Sandbox, Functions), similar to Next.js's early lock-in dynamics with Vercel.