Vercel launches Drives: agents get persistent storage of up to 16 TiB
Vercel has announced general availability of Vercel Drives, persistent disks for Vercel Sandbox. An agent's files now live independently of the runtime environment. A new sandbox with different tools, a different runtime or more compute can mount the same disk and pick up where the last one left off. The feature is available on all plans. The default size is 1 TiB (1 GiB on Hobby), with a maximum of 16 TiB. • The disk is mounted with the sandbox create --mount command or through the SDK • Other sandboxes can attach it in read-only mode • Disk forks give each agent its own copy; they do not merge back into the original
AI-processed from Vercel Blog; edited by Hamidun News
Vercel has announced general availability of Vercel Drives, persistent disks for Vercel Sandbox. The idea is that an agent's files are no longer tied to the sandbox it runs in. You create a disk once and mount it in any new environment. The feature is available on all plans: the default disk size is 1 TiB (1 GiB on Hobby), and the maximum goes up to 16 TiB.
Files separate, environment separate
Agents often do complex work that stretches across several runs. Requirements change along the way: the next stage may need a different set of tools, a different runtime or more compute. If the working directory lives inside the sandbox, changing environments means moving the files. Drives removes that step.
Setup takes two CLI commands:
sandbox drives get-or-create agent-workspace sandbox create --mount agent-workspace:/workspace
In the SDK, the same thing is done with Drive.getOrCreate and the mounts parameter of Sandbox.create. In Vercel's example, the agent clones a repository with git clone straight into /workspace/repo. Everything under /workspace is saved on the disk. Everything else belongs to the sandbox and disappears with it. The next sandbox that mounts this disk will find the repository and any other agent files exactly where the agent left them.
Shared data and forks
A disk can also serve as shared storage. One agent prepares a dataset, and others mount copies of it in read-only mode. They analyze the data or build reports without rebuilding the dataset from scratch, and they write their results separately, leaving the source files untouched.
There is a second scenario: disk forks. Each fork starts with the contents of the source disk and gets its own writable workspace. Several agents can try different transformations of the same dataset in parallel without getting in each other's way and without changing the original. Forks do not merge back into the source disk, so choosing the best result and moving it where it needs to go is left to the developer or the agent itself. Forks have also reached general availability.
What's in the release
- Status: general availability for Vercel Sandbox.
- Plans: all, including Hobby.
- Default size: 1 TiB, 1 GiB on Hobby.
- Maximum size: up to 16 TiB.
- Mounting: CLI (the --mount flag) or Sandbox SDK (the mounts parameter).
- Sharing: one disk can be mounted from other sandboxes, including read-only.
- Forks: independent writable copies of a disk, with no merging back.
- Also: snapshots and limits are described in the documentation, and the GA changes in the changelog.
How to get started
The easiest approach is to take an existing agent workflow and move its working directory onto a disk: create it with get-or-create and mount it at /workspace. This draws a clear boundary: whatever must outlive a run sits on the disk, while temporary artifacts and caches stay in the sandbox. For long multi-stage tasks, it makes sense to use one environment for assembling data and another, with a different runtime, for analysis. You won't have to move files between them.
Forks suit experiments: one prepared dataset, several independent processing branches. For bulk analysis, it's better to hand out read-only copies so that an accidental write can't corrupt the source data. Before running in production, check the documentation on limits and snapshots, especially if you need sizes above the default. For a quick start, Vercel offers a starter template, Vercel Drives eve agent, which deploys in one click.
Delegate Deep Research, Coding & Workflows to an AI Agent
Deploy autonomous multi-step execution with NomadNet Agent (agent.nomadnet.ai) — automated web research, data synthesis, code generation, and complex task orchestration.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.