Vercel brings persistent agent storage out of beta across all plans

An agent can move to a new computing environment without leaving its workspace behind—but only files saved on its mounted drive make that trip. Vercel made Drives generally available across all Vercel Sandbox plans on October 15, 2026, separating persistent files from the environment running the agent rather than preserving everything inside that environment.

This is a move out of beta, not the first availability of persistent storage. Vercel’s September 23 public-beta announcement already offered Drives on Hobby, Pro and Enterprise plans. That release described reusable storage for agent workspaces, files used as memory, datasets, models and installed software dependencies.

The setup starts with a drive that exists independently of a sandbox—the environment where an agent executes its work. A developer creates or retrieves the drive, then mounts it at a directory when creating a sandbox. Files written under that directory stay on the drive; the rest of the filesystem belongs to the sandbox.

In Vercel’s example, the agent downloads a code repository into /workspace/repo. The next sandbox mounting that drive finds the repository and other saved files where the previous run left them. The agent can resume in a fresh sandbox or switch runtimes—the software used to run its code—without transferring that workspace.

For work spread across agents, Vercel describes two different paths. One agent can prepare a dataset on a drive, while others mount read-only copies to analyze it or produce reports. Those agents use the prepared files as input and save their results separately, rather than rebuilding the dataset for each task.

The beta announcement explained how shared reading works: multiple sandboxes can mount point-in-time, read-only snapshots after a drive has been written to. Each snapshot reflects the drive when mounted. Later writes are not included, so a sandbox needs a newly mounted snapshot to read those changes. The beta also specified one read-write mount at a time.

Forking, now also generally available, serves a different purpose. Each fork begins with the original drive’s contents but supplies its own writable workspace. Agents can try different dataset transformations in parallel without changing the original or affecting one another.

Drives default to 1 TiB, compared with 1 GiB on Hobby, and can be configured up to 16 TiB. Those capacities are unchanged from the public beta.

The beta’s deployment rules allowed each sandbox to mount up to four drives at separate paths. A drive stayed in the region where it was created, and any sandbox mounting it had to run in that region without failover regions.