Google launches open-source Agent Substrate on GKE

Google has made Agent Substrate available on Google Kubernetes Engine, positioning it as an open-source runtime for running large numbers of AI agent sandboxes.

The software is aimed at teams moving from local AI agent testing to large-scale deployments involving hundreds of thousands of concurrent agents. Those environments create pressure around isolation, network controls, scheduling, and the cost of idle compute capacity.

Agent Substrate runs on Kubernetes infrastructure and is tuned for Google Kubernetes Engine, while remaining available as open-source software for other Kubernetes environments. It is designed to separate agent execution from machine management, with Kubernetes handling the underlying machines and a separate execution layer managing sandbox lifecycles.

Users can choose between Cloud Hypervisor microVMs and gVisor sandboxes for isolation. The platform also includes an integrated gateway to manage inbound and outbound traffic and apply network controls around agent activity.

Google said the software is built for workloads in which AI systems generate and run code, interact with command-line tools, use browsers, and work with files over long periods. In those cases, conventional container approaches can struggle when many agents need to be created, paused, and resumed in quick succession.

According to Google, Agent Substrate can resume sandboxed environments in less than 500 milliseconds and handle more than 500 suspend-and-resume activations per second. It can also support more than 1,000 dormant agents per host by suspending idle sessions and releasing CPU and memory until work resumes.

Scale focus

A central part of the launch is the claim that AI agent infrastructure behaves differently from standard application infrastructure. Autonomous agents often spend much of their time waiting for model inference, tool responses, or human input, making permanently allocated resources inefficient, Google said.

To address that, the software snapshots agent state to local disk and Google Cloud Storage. Those snapshots allow sessions to return to a previous state without keeping full compute resources tied up while the agent is inactive.

The launch also reflects a broader effort by cloud providers to shape infrastructure around AI agents rather than only model training and inference. Businesses experimenting with coding agents, command-line assistants, and automated workflows are beginning to seek systems that can isolate untrusted code while keeping operating costs under control.

Google said Agent Substrate is intended to work with a range of agent frameworks and harnesses, including Claude Code, OpenClaw, and Hermes. It also named Antigravity and Codex among tools developers already run locally before attempting to scale them across much larger fleets.

Cloud tie-in

Although the software is open source, Google highlighted several features tied to its own cloud infrastructure. These include custom ComputeClasses in Google Kubernetes Engine to manage machine pools across different processor types and pricing models, including spot and on-demand capacity.

Google also pointed to support for Google Axion, its Arm-based processor line, saying this can improve price-performance for sandbox workloads compared with rival cloud offerings. For stateful workspaces, Agent Substrate can be integrated with Filestore agent volumes so network file system mounts can be attached and detached quickly.

That approach targets workloads where agents need shared file access across multiple turns or between several agents. Google said the storage option supports read-write-many access and POSIX-compliant file locking to reduce the risk of write conflicts during collaborative work.

Early user

Among the first users Google cited is Nous Research, developer of the Hermes agent. Google said Hermes ranks first globally by OpenRouter usage across productivity, coding, command-line, and personal agent categories.

According to Google, Nous Research has worked with the company as an early design partner while assessing how the runtime manages agent isolation and identity requirements.

"We built Hermes Enterprise to enable customers to deploy into their existing infrastructure, while handling per-agent isolation and extensible access control. Agent Substrate addresses both at the platform layer in a way that also preserves valuable compute resources. Our experience with Agent Substrate gives us confidence the architecture can scale efficiently as agent workloads grow," said Hervé Bizira, Chief Business Officer, Nous Research.

Agent Substrate is available to Google Kubernetes Engine customers for non-production workloads, while Google said it is offering general availability support for production deployments through an allowlist process.