Vercel has introduced Vercel Managed Images for Sandbox, changing how developers configure isolated Linux environments for AI agents, code generation tools, and other workloads that need to run untrusted code.

The company says the new images are versioned, open source base images that can be used directly or extended. Their source lives in the public vercel/sandbox repository. Managed images replace Vercel's older Sandbox runtimes, which the company now describes as deprecated.

The default path also changes with version 3 of the Sandbox SDK. New sandboxes now default to vercel/sandbox/universal:latest, an Ubuntu-based image that includes Node.js, Python, common coding agents, and standard utilities. Vercel says this should reduce the need for developers to build custom images or install packages each time a sandbox starts.

For production agent systems, the operational details matter more than the branding. Vercel says each managed image gets a nightly release, while rolling tags such as latest and major-version tags pick up operating system updates, dependency updates, and security patches automatically. Within a given release, dependencies are pinned where possible so teams can use fixed image versions when they need repeatable environments.

Vercel's Sandbox documentation describes the underlying product as a compute primitive for running arbitrary code in isolated, ephemeral Linux VMs. It supports managed images, custom OCI images stored in Vercel Container Registry, Firecracker microVM isolation, and per-agent isolation patterns. The managed-image shift is therefore a developer-infrastructure update: less about a new model, and more about making agent execution environments easier to patch, reproduce, and share.