General Availability of Untrusted Code Execution via Vercel Sandbox

Unlike traditional human-centric infrastructure, AI agents repeatedly execute code quickly in a secure, isolated environment and then immediately delete it.

Vercel has launched the general availability of Vercel Sandbox, an execution layer designed for this characteristic of agents. This service leverages a Firecracker-based microVM cluster to provide infrastructure that can create and dispose of isolated environments within seconds.

The core features of Vercel Sandbox are as follows:

• Ultra-fast startup: Runs thousands of sandboxes per task in under 1 second
• Complete isolation: Protects systems from untrusted code in user input and repositories
• Ephemeral environments: An ephemeral structure that exists only as long as needed and shuts down automatically
• Snapshot capability: Reduces time by instantly restoring environments via snapshots instead of rebuilding them
• Efficient billing: Cost optimization based on Active CPU usage rather than idle time

Vercel has open-sourced a CLI and SDK so developers can take advantage of this infrastructure. Services such as Roo Code and Blackbox AI are currently adopting it to maximize the efficiency of their agents' work.