Tensorlake

Special Edition · August 2026

Agent state that outlives the sandbox

Introducing Tensorlake Cloud Volumes and Git Repositories: a shared drive your agents can mount anywhere, and real Git with no clone step that protects their work before it's ever published.

In this special edition. We're excited to announce two new features: Tensorlake Cloud Volumes and Git Repositories. With these, you no longer have to worry about your agents' work getting lost when sandboxes come and go. Cloud Volumes give an agent a shared folder that mounts anywhere, on any sandbox provider or cloud, and automatically backs itself up as your agents work. Git Repositories connect to a sandbox in about a second, no clone needed, and bring real version control, agents can work in private and you decide when to publish the changes. The server handles merges and keeps a record of what every agent did. Start with the Cloud Volumes docs or the Git Repositories docs.

Cloud Volumes

A directory that saves and versions itself. Mount it anywhere. Rewind anytime.

A Cloud Volume shows up in your sandbox as a regular folder. Your agent can read and write files the same way it always does, and every change is automatically backed up to durable cloud storage within about a second. If the sandbox crashes halfway through, just mount(attach) the volume to a new one and your agent can keep going.

Take a snapshot to create a checkpoint that you can always roll back to, or fork a copy to try something different without touching the original. You can have many agents sharing the same volume, and changes to different files merge automatically. And volumes are fully portable: you can use them with any sandbox provider, whether that's AWS, GCP, Azure, Kubernetes, or even your laptop.

Start with the Cloud Volumes docs →

Git Repositories

Real Git, rebuilt for agents. No clone, attach in a second.

Attaching a repository to a sandbox takes about a second, no matter how large it is, because there is no clone step, files download only when your agent reads them. That makes fan-out cheap: mount the same repo into a hundred sandboxes and the hundredth attaches as fast as the first one. Tensorlake repositories are still plain Git, so clone, push, and merge just work like you expect. Under the hood, we rebuilt it for agents, so it can handle tens of millions of repositories and hundreds of commits per second on a single repo.

Why not just run git in the sandbox? Because regular Git only saves what's already pushed. If an agent crashes halfway through, you lose everything else, unless you are constantly committing and pushing just in case. Tensorlake mounts fix that: agents work in private by default, each with its own draft space and autosave running every 30 seconds without creating a single commit. That draft space lives on Tensorlake's servers, not in the sandbox, so if the sandbox dies, mount the repo in a new one and the agent picks up from its last autosave. When something is ready, you promote it into the shared history, with a full record of which agent did what. Merges and conflict checks happen there too, so sandboxes never have to download a whole repository just to combine changes. And nothing gets force-overwritten: if there's a conflict, nothing publishes and you get a report of exactly which files collided.

Start with the Git Repositories docs →

Check out our docs to get started

Tensorlake · AI Infrastructure for Agents