Google Agent Compute Engine: A MicroVM Per Agent, Plus GKE

Google says it learned from scaling Gemini that running autonomous agents safely is “fundamentally an infrastructure challenge,” and it is now selling the answer. Agent Compute Engine, the service that already gives every Gemini agent its own microVM, is coming to customers in preview in the coming months. For platform teams choosing where agent code runs, the bigger near-term news may be its open-source sibling for Kubernetes, which reports as much as 10x higher compute density.

What Happened

At AGNTCon + MCPCon, Google Cloud announced three things:

• Agent Compute Engine: a compute service built for agents that write and run code, call tools and drive browsers. Preview is due in the coming months, with an interest form open now.
• Agent Substrate on GKE: an open-source execution runtime that brings the same agent-first design to any Kubernetes cluster, available to try today.
• Wider governance: Gemini for business agent controls extended across more runtimes, other clouds and local devices.

Why It Matters

Agents have an unusual lifecycle. They start, pause and resume constantly, and much of a session is spent “thinking” between tool calls. Agent Compute Engine runs each agent in a fast-booting microVM. It shares memory copy-on-write across agents built from one OS template, and pauses idle microVMs with checkpoint and restore during the thinking phase or while an agent waits on a person or an external service.

Isolation is the second design goal. Each microVM sits behind a hardware-isolated Cloud Hypervisor boundary with its own guest kernel, runs on Google’s Titanium offload hardware, and uses its own agent identity with short-lived credentials. Each also gets its own Hyperdisk so workspaces and agent context persist across long sessions. Google’s technical deep dive covers the architecture in detail.

Agent Substrate makes the same approach portable. When an agent goes idle, it snapshots memory, returns the machine to the pool and resumes on demand. Google’s figures:

It supports microVM or gVisor isolation, gives each agent a cryptographic identity, injects credentials through egress proxies outside the agent’s reach, and works with harnesses including Claude Code, OpenClaw and Hermes.

What This Means for Platform Teams

Governance is spreading beyond Google’s own runtimes. Agent Gateway, Agent Registry, SPIFFE-based Agent Identity and OTel-compliant agent observability are being extended across Agent Runtime, Cloud Run, GKE and Agent Compute Engine. Envoy-based enforcement for other clouds and on-prem, and endpoint governance for coding agents like Claude Code, are due in preview soon. Token budgets and model routing policies for Agent Gateway arrive in preview this quarter.

Stack Archive compared how AWS, Azure and Google Cloud built agent sandboxes; this launch adds a dedicated compute service on top of those sandboxes.

Source: Google Cloud Blog

Content created with AI assistance and reviewed for accuracy.