Vercel launches Run SDK for secure agent code execution

Vercel released Run SDK, an open-source package for executing untrusted JavaScript and TypeScript inside a sandboxed QuickJS context — purpose-built for agents that generate and run code.

What changed. Run SDK runs user- or agent-generated code in a worker thread with no direct access to Node.js, the filesystem, or the network. Applications expose only narrow hostFunctions that the sandboxed code can call. It supports human-in-the-loop interruptions for sensitive operations (like approvals or auth), signed tokens to resume without replaying completed work, and configurable limits on heap size, function call depth, and output size. The builder is already powering code mode in the AI SDK and agent-generated TypeScript in Eve. It requires Node.js 22.13+ or Bun.

Why a builder cares. Every agent framework that writes code eventually needs a sandbox — not a Docker container per request, not eval(). Run SDK is a focused, embeddable alternative that ships as a single npm dependency. If you're building an agent that needs to safely run generated TypeScript against curated tool calls — with pause-and-resume for approvals — this is the production-grade primitive.