Agent workspaces need their own storage lifecycle — persistent state is becoming infrastructure

Google Cloud introduced Filestore agent volumes on September 15, 2026 as a managed storage layer built for stateful agent sandboxes, autonomous fleets and multi-agent developer environments. It is a useful signal that agent infrastructure is expanding beyond models, prompts and tools into the filesystem itself.

Filestore agent volumes provide isolated, POSIX-compliant persistent storage that can be mounted into agent sandboxes. Google positions the capability around fast workspace provisioning, state persistence across suspend-and-resume cycles, and shared read-write access when multiple agents need to work on the same project files.

The product is available broadly for evaluation and non-production use, with production support offered through a limited allowlist. That matters because the design pattern is visible even before the service reaches universal production maturity.

Long-running agents increasingly create artifacts, install packages, modify repositories and accumulate intermediate state. Treating all of that as prompt context is inefficient, while tying state to a disposable container makes resumption fragile.

A durable workspace lets compute scale independently. The platform can suspend an idle agent, retain its files, and reconnect later without paying continuously for the same compute footprint. Multi-agent work also becomes easier when collaboration happens through a controlled shared filesystem rather than by copying large blobs through model context.

Outborn AI can treat workspace state as a separate runtime resource with its own lifecycle and retention policy. Outborn Agent Connect can govern which external systems are allowed to materialize files into that workspace. AppKit can register workspace ownership, environment and storage policy. Obsi Drive and Obsi Knowledgebase can become durable user-facing destinations for approved outputs rather than temporary scratch space.

The boundary is important: scratch files, durable records and governed knowledge are not the same thing. Outborn products should make that distinction explicit.

• Separate disposable compute from persistent agent workspace state.
• Assign workspaces to a user, agent, application and environment.
• Make retention and deletion policy explicit.
• Prevent one agent session from silently inheriting another tenant’s files.
• Use controlled shared workspaces for multi-agent collaboration.
• Promote approved outputs into Drive or Knowledgebase rather than keeping everything in scratch storage.
• Preserve file provenance so consequential outputs can be traced back to their generating session.

Agent storage will likely become a common platform layer alongside sandboxes and durable sessions. The design challenge will be less about raw capacity and more about isolation, cost, lifecycle policy and secure sharing across humans and agents.

• Google Cloud — Introducing Filestore agent volumes
• Google Cloud — About Filestore agent volumes

Outborn AI · Outborn Agent Connect · AppKit · Outborn Governance · Obsi Drive · Obsi Knowledgebase

Workspace disks hold more than convenience: cached inputs, generated artifacts, dependency trees and potentially sensitive credentials. Retention, backup, encryption and ownership must therefore be explicit policies rather than accidental consequences of VM reuse.

Terminate an agent mid-task and restore its workspace. Verify that approved artifacts survive while temporary secrets and unintended private data do not.

Conceptual workflow, not an operational assertion about a vendor.

Measure restore correctness, cross-tenant isolation, retention expiration and orphaned storage cleanup.

Reference: Primary technical guidance. This analysis is a design interpretation, not an independently executed product benchmark.