Google is turning the infrastructure that runs Gemini into a product. At AGNTCon + MCPCon on October 22, Google Cloud launched Agent Compute Engine, a purpose-built compute service that gives every AI agent its own hardware-isolated micro virtual machine. The service already powers Gemini’s agent sessions, Google says, and will open to customers in preview in the coming months so they can run their own custom agents the same way.
Alongside the launch, Google open-sourced Agent Substrate, a Kubernetes runtime built on the same design, and extended its Gemini for business agent governance tools — identity, registry, gateway and observability — to agents running on other Google Cloud runtimes, other clouds, on-premises systems and local devices.
Agent Compute Engine: A MicroVM For Every Agent
Agent Compute Engine was built for the way agents actually work: writing and executing code, calling tools, driving browsers and looping in short bursts, rather than running steadily like a normal workload. Each agent gets a microVM with its own guest kernel and root access, behind a hardware-isolated Cloud Hypervisor boundary. The isolation runs on Titanium, Google’s offload infrastructure processing unit, which Google describes as silicon-rooted defence for running untrusted agent code. Every microVM carries its agent’s own identity with short-lived credentials, gets a cryptographically verified ID, and receives service credentials injected in egress, outside the microVM itself. Each one also has its own Hyperdisk, so an agent’s workspace and context persist across long-running sessions.
Cost is the other half of the design, as The Decoder noted in its coverage. Agents spend much of their time waiting — on a tool, a human or another system — so Agent Compute Engine pauses idle microVMs by fast checkpoint and restore, and shares memory between VMs through copy-on-write from a single operating system template. When an agent finishes a tool call and enters its thinking phase, its microVM signals idle and the host parks it until it is needed again.
Agent Substrate: The Same Design, Open Source On Kubernetes
For platform teams that want portability, Agent Substrate brings the same agent-first design to Google Kubernetes Engine and, Google says, to any Kubernetes cluster in any environment, with any agent framework — it names Claude Code, OpenClaw and Hermes. Google claims as much as 10x higher compute density than traditional compute while keeping full kernel isolation, and says that in benchmark validation running 200,000 agents, Agent Substrate sustained 500 activations per second at less than 500 milliseconds of suspend and resume latency. Nous Research is already using the architecture for its Hermes Enterprise deployments, its chief business officer Hervé Bizira said. The approach parallels the sandbox work covered in VAST Data and NVIDIA’s agent sandbox stack, and Google’s own earlier research into agents that rewrite their own harness: isolation and density are becoming the contest, not just model quality.
Governance Extends To Agents Wherever They Run
The governance half of the announcement widens the reach of tools Google already ships in its Gemini platform: Agent Gateway, which routes and polices agent traffic; Agent Registry, the catalogue of approved agents, tools and models; Agent Identity, which gives each agent a SPIFFE-based secure ID; and OpenTelemetry-based observability for every model call and tool invocation. Those controls now extend across Agent Runtime, Cloud Run, GKE and Agent Compute Engine, with Envoy-based enforcement points for multicloud and hybrid setups in preview soon, and OS-level sandbox governance for local coding agents such as Claude Code and Antigravity CLI on managed devices.
New Gateway features arriving in preview this quarter include granular token and cost controls, model routing policies, and delegated authorisation through third-party identity providers Ping Identity, SailPoint and Saviynt — the outside identity support The Decoder highlighted. L’Oréal’s AI lead engineer Alexandre Frizzo said Google Cloud’s governance gives the company a foundation to scale its Beauty Tech Agentic Platform with control over what each agent can access.
Why It Matters
Agent Compute Engine is Google pricing a problem every agent platform now has: an agent that waits is still a computer that costs money, and an agent that acts is a security risk that needs its own walls. By pausing idle microVMs and sharing memory, Google attacks the cost side; by giving every agent its own kernel, identity and credentials held outside its reach, it attacks the risk side — the same risk side that NVIDIA’s OpenShell architecture targets from the hardware up. If the preview lands as described, the infrastructure Gemini runs on becomes a template rivals have to match: agents sold not as a model with tools attached, but as governed, isolated computers that happen to think.
Background
Google has been rebuilding its cloud around agents through 2026, from the Gemini platform’s Agent Gateway and Registry to the open-sourcing of Harness2, its framework for agents that improve their own scaffolding. Agent Compute Engine comes out of a specific lesson Google says scaling Gemini taught it: single-turn prompting needed only ordinary compute, but interactive agents that execute code and browse the web need execution environments designed for their stop-start, high-churn lifecycle — and governance designed for fleets of them, wherever they run.
Source: Google Cloud Blog; The Decoder
0 Comments