Agent files stop dying with the sandbox: Vercel Drives hits general availability, so a new environment resumes the same repo

An agent's working files can now exist independently of the sandbox running it, so work continues in a new environment without transferring the workspace. Previously an agent's files were tied to its execution environment, and moving across runs or runtimes required transferring the workspace. The documented usage creates a drive once and mounts it when creating a sandbox; the next sandbox mounting that drive finds the repository and written content from the previous run.

Persistence is scoped to the mounted path, while everything outside that path still belongs to the sandbox. Durability is decided by the mount point rather than by retaining the whole sandbox state. In the example the drive is mounted at /workspace, everything under it persists on the drive, everything else belongs to the sandbox, and other sandboxes can mount the same drive.

One drive can be shared across sandboxes, or used in parallel through read-only copies and forks. Sharing and parallel exploration no longer require rebuilding the dataset, and they leave the original contents unchanged. After one agent builds a dataset on a drive, other agents can mount read-only copies to analyze it or generate reports; a fork starts from the drive's contents with its own writable workspace and does not merge back.

The capability targets developers running agents in Vercel Sandbox and is generally available on all plans. Drives default to 1 TiB, 1 GiB on Hobby, with up to 16 TiB available; configuration, snapshots and limits are documented officially.

Forks do not merge back into the original drive, so results from parallel exploration must be consolidated separately; when read-only copies are shared, each agent's results are saved separately. Specific capacity and snapshot limits are described in the documentation.