Google Cloud launches Agent Compute Engine, open-sources Agent Substrate for hardware-isolated AI agents
Google Cloud announced Agent Compute Engine (ACE) at AGNTCon + MCPCon on October 22, 2026, a purpose-built compute service for AI agents that already powers Gemini as an agent for work tasks. Google says ACE will be available in preview in the coming months, providing hardware-isolated runtime sandbox security, and paired the launch with expanded agent governance in Gemini for business.
ACE gives each AI agent a hardware-isolated Cloud Hypervisor microVM with its own guest Linux kernel and root access. Every microVM receives a cryptographically verified identity, and service credentials are injected at egress outside the microVM, keeping them out of reach of the agent itself while enforcing granular network policies.
To support multi-day tasks without losing context, each microVM integrates with Google Cloud Hyperdisk, decoupling ephemeral compute from durable state across a three-tier storage model: fast disposable overlays for operating system churn, shared read-only toolchains mounted via FUSE on Google Cloud Storage, and persistent user workspaces. The design targets sub-second resume latency so suspended agents can scale alongside active compute.
For self-hosted deployments on Google Kubernetes Engine, Google open-sourced Agent Substrate, an execution runtime that brings the same agent-first design to Kubernetes. In benchmark validation running 200,000 agents, Agent Substrate sustained 500 activations per second at under 500 milliseconds of suspend and resume latency, achieving up to 10-times higher compute density than traditional compute while maintaining full kernel isolation. It supports agent harnesses including Claude Code, OpenClaw and Hermes.
On governance, Google is extending Gemini for business controls to support agents wherever they operate, including local coding agents such as Claude Code and Antigravity CLI on managed devices through an OS-level sandbox and local governance services, which Google says are coming to preview soon.
For enterprise platform teams, the launch is a production-grade blueprint for running untrusted agent-generated code safely, but buyers should weigh the caveats: the managed ACE service is not yet in public preview, Google has published no commercial pricing for microVM reservations or storage tiers, the 500-activations-per-second figure is synthetic validation rather than a live enterprise workload, and credential injection outside the microVM will need alignment with existing IAM and service perimeter policies.