Google Cloud launches Filestore agent volumes for AI

Google Cloud has launched Filestore agent volumes for AI agent workloads. The service is available to Google Cloud customers for non-production use.

The product adds a managed file storage option for companies running large numbers of software agents in isolated sandbox environments on Google Kubernetes Engine. It is designed for tasks such as coding, data analysis and collaborative research, where agents need persistent workspaces that can be created and removed quickly.

Platform teams have often met those storage needs with custom systems, moving state between temporary agent environments and central storage or managing local disks and self-hosted file systems. That approach becomes harder to manage as the number of concurrent agent sessions rises into the thousands or millions.

Filestore agent volumes are tied to Google Cloud Filestore and work with Agent Substrate on GKE and GKE Agent Sandbox. The storage lifecycle is automated, so when a new sandbox starts for an agent task, a dedicated file workspace is allocated and attached in milliseconds.

This is intended to reduce the need for engineers to create, attach and remove storage volumes for each task manually. It also addresses a common trade-off in agent infrastructure between keeping storage ready for fast start-up times and limiting the cost of unused capacity.

Storage model

The new volumes are built around isolated persistent workspaces for each agent environment. That separation is intended to limit security risks and data leakage when agents run dynamic code, install third-party packages or use external tools.

The service also supports shared access across multiple agents through read-write-many access and POSIX file locking. This allows orchestrated groups of agents to work in the same file tree at the same time while reducing the risk of write conflicts.

Google is targeting several use cases. One is software engineering, where agents may need to install libraries, write patches, run build tools and carry out unit tests inside separate workspaces. Another is longer-running workflows that can be paused when idle and resumed later without rebuilding the working environment from scratch.

A third is multi-agent collaboration. In those arrangements, a lead agent might assign tasks to research, code generation and validation agents, all working from the same project files.

Cost focus

The service uses a pay-per-use model based on storage capacity consumed, with automatic lifecycle tiering for inactive workspace data. This is intended to cut waste for workloads in which many agent sessions are short-lived or intermittent.

For cloud customers, the launch reflects a broader shift in infrastructure design as interest in autonomous and semi-autonomous software agents grows. While much of the market's attention has focused on models and compute, storage has become a practical issue for operators trying to run agent systems at scale without high idle costs or slow restarts.

Google has been building out products around what it describes as agentic workloads, particularly on Kubernetes-based infrastructure. By linking the new storage option to GKE services already used for isolated execution, it is positioning Filestore as part of a broader stack for teams developing agent platforms on its cloud.

Production support for Filestore agent volumes is being offered through an allowlist. General access for standard production workloads has not yet opened more widely.

Google says Filestore agent volumes are intended for organisations building agent-as-a-service platforms, coding assistants and larger internal agent fleets. Each time GKE launches a sandbox for a new agent task, Filestore automatically allocates and attaches a dedicated, isolated file workspace to that environment in milliseconds.