Skip to main content
Agent Workspace is the durable logical working state available to an instance across sessions, whether implemented as a persistent filesystem, database-backed state, object storage, or retained session state.
A later run can use and change the state left by an earlier run. It continues from the work itself instead of reconstructing it from a transcript. A simple test is to open another session a day later. If the agent can continue the earlier work, it has a durable computational workspace. The term describes a logical contract. A platform may use databases or object storage. It can also give each session an ephemeral filesystem, retain selected changes, and make them available to later sessions. The paper’s reference design gives every instance a persistent virtual computer: a filesystem, a shell, and network access, wrapped by policy code. The agent can keep documents, code, local databases, installed packages, artifacts, and unfinished work there without first defining a schema for each kind of state. The platform can snapshot and restore that environment, so an idle instance does not need a machine to keep running. The workspace is also the privacy boundary. It includes every file, database record, object, and session artifact the instance can retrieve. Information that must remain separate belongs in separate workspaces. Authentication controls who may use the instance. Separate privacy compartments require separate workspaces. Plugin package content belongs to the immutable release. Writable plugin data belongs to the instance workspace and is scoped by application, instance, and plugin identity. It follows the same isolation, retention, export, checkpoint, and deletion rules. Credentials remain in a managed secret store. A context window ends with the model call. The workspace remains available to later sessions. See the paper’s category test and workspace architecture.