Faster agent workspaces: content-addressed file storage

Every conversation with a specialist has a workspace: the files it keeps between turns. We used to save that folder as a single archive, so every cold start downloaded, unpacked, and later re-uploaded the whole thing. That scaled badly: one design conversation with 175 files and 260 MB of images saw its sandbox start climb from 2 to 15 seconds to more than 200 before it failed outright. Workspaces are now stored the way Git stores a repository, as content-addressed objects plus a manifest, synced incrementally.

How it works

• Stored once. Each file is stored once by its content hash, and small files are grouped into packs, so a project with thousands of tiny files doesn't become thousands of requests. Unchanged files are never uploaded twice.
• No lost work. A new manifest is published only if the previous one is still current, so a late or stale save can't overwrite newer work or bring back deleted files.
• Restores straight to the sandbox. Files over 25 MiB upload from the sandbox itself instead of passing through our servers.
• Migrated automatically. Older workspaces convert the next time they're used.

Also shipped

• When a workspace grows past 25 MiB in one file or 128 MiB in total, the agent sees the largest files and is guided to archive inactive ones to storage you've connected.
• Vercel joined E2B as a supported sandbox provider for deployments that configure it.