rt.workspaces manages the persistent storage you mount into sandboxes and hand to agents as filesystem memory (see Workspaces for the concept). The bytes inside a workspace live under rt.workspaces.files.
Manage workspaces
create() provisions the backing storage root; if the org is over its storage quota, it fails before the workspace is usable. delete() returns 409 WORKSPACE_IN_USE when a sandbox, agent, or installable agent still references the workspace — check usage() (and the REST usage endpoint) first, or force from the API.
Files
rt.workspaces.files reads and writes the contents of a workspace. Paths are workspace-relative.
Only organization workspaces can be mounted into sandboxes; personal workspaces are reserved for thread storage and personal files. With a scoped API key, reads need
files:read and writes need files:write.Mount into a sandbox
Mounting is what turns a workspace into an agent’s durable memory:/workspace outlives the sandbox and reappears in any sandbox that mounts the same workspace. See Workspaces for the full memory model and the Workspace Files API for the REST surface.