Google Cloud gives every AI agent its own microVM with Agent Compute Engine
Google Cloud has introduced Agent Compute Engine, a compute service that gives each AI agent its own hardware-isolated microVM with zero stored credentials, alongside open-source sibling Agent Substrate for Kubernetes.
Google Cloud has introduced Agent Compute Engine, an infrastructure service built to host AI agents, unveiled at AGNTCon + MCPCon North America in San Jose (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/, https://aidailypost.com/news/google-cloud-launches-agent-compute-engine). The core promise: one hardware-isolated microVM per agent — own guest kernel, root access, Hyperdisk storage — with zero credentials inside the sandbox.
The design splits every agent in two. The “harness” — the brain — calls the model and runs entirely outside the sandbox, while the “sandbox” — the hands — executes code, shell commands and browser actions inside the microVM (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/). Credentials stay outside; Google’s Agent Gateway submits them on the agent’s behalf, so a prompt-injected agent finds no secrets to steal (https://siliconangle.com/2026/10/22/google-cloud-introduces-agent-compute-engine-to-host-ai-agents/).
The microVMs run on Google-managed Titan-silicon hosts with the Titanium IPU, behind a Cloud Hypervisor boundary; idle agents are checkpointed and resumed from snapshots. Google says the service already powers Gemini as an agent for work tasks, plus Antigravity and internal evaluation and reinforcement-learning pipelines (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/).
Google also released open-source sibling Agent Substrate for any Kubernetes cluster, with support for frameworks including Claude Code, OpenClaw and Hermes. Google claims 500 activations/second, suspend/resume under 500 milliseconds at 200,000 agents, and up to 10x density over traditional compute (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/).
New governance controls — token tracking, model routing, delegated authorisation via Ping Identity and SailPoint — are arriving in stages across Cloud Run, GKE and managed endpoints (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/). Customer preview is targeted for “the coming months”; Google has published no pricing, boot times, density figures or firm date (https://www.marktechpost.com/2026/10/22/google-cloud-agent-compute-engine-microvm-ai-agents/).
More on this topic: all Technology stories


