Google is turning the infrastructure that runs its own Gemini agents into a product. At the Linux Foundation's AGNTCon + MCPCon in San Jose on Thursday, Oct. 22, 2026, Google announced Agent Compute Engine, a compute service that gives every AI agent its own hardware-isolated microVM. Google said in a blog post that it will be available in preview "in the coming months."
What it does
- A computer per agent: each agent gets a persistent, sandboxed machine where it can run shell scripts, drive a headless browser and write code with root access, in loops that could run for weeks, The New Stack reported.
- Pause while thinking: when an agent finishes a tool call and waits on the model or a person, the microVM is checkpointed and paused, then resumed from a snapshot, Google said.
- Nothing to steal: sandboxes hold no secrets. Credentials are added outside the microVM at the network egress, and each one has its own short-lived identity, according to Google. It runs on Cloud Hypervisor with Google's Titanium offload hardware.
- Already in use: it started as internal infrastructure for Gemini, Antigravity and Google's model-training pipelines, The New Stack said.
Google also extended its Gemini for business agent-governance controls to agents running in other places, and pointed Kubernetes users to Agent Substrate, its open-source runtime for GKE. Google says Agent Substrate handled 500 activations per second in a 200,000-agent benchmark, The New Stack reported.
The competition
Agent sandboxes aren't new. The New Stack noted that startups E2B and Daytona offer them, and AWS's Bedrock AgentCore also gives each agent session its own microVM. The race to give AI agents safe computers to work on is the same one behind this week's launch of Anthropic's Claude Duet and xAI's Grok Bot pages.