Lazy-loading agent workspaces with a real filesystem

Every conversation gives its specialist a persistent work folder in its sandbox. Until now, a cold sandbox fetched that folder's contents before the agent could act, so the bigger the workspace, the longer the wait. As of October 8, the work folder is a real filesystem, built on open-source JuiceFS, that restores only names and metadata when the sandbox starts and fetches each file's data the first time something reads it.

How it works

• Lazy restore. Metadata is restored with the sandbox, file data is fetched from storage on first access and cached, and small files are prefetched in the background so a full scan doesn't pay a round trip per file.
• As fast as a native sandbox. A restore is ready in about the time a fresh sandbox takes to boot, for workspaces up to 5,000 files and 128 MiB.
• Real filesystem behavior. Large files, random reads, memory mapping, permissions, and links behave as on a local disk, so shell and file tools need no changes.
• Safe checkpoints. Checkpoints run periodically and when a run pauses, completes, or hands off; a failed checkpoint keeps the previous one and retries.
• Isolation. A sandbox never holds our storage keys. It gets credentials for its own workspace only, valid for at most 24 hours, and a storage gateway checks every request.
• Limits. A workspace holds roughly 190,000 files; past that, the agent is told to remove rebuildable folders like node_modules.