Google Cloud announced today, October 22, 2026, Agent Compute Engine, a compute service purpose-built for artificial intelligence agents. The platform is already used internally to power Gemini agents and is expected to enter preview for customers in the coming months.
What happened?
The announcement of Agent Compute Engine came during the AGNTCon and MCPCon events. According to Google, the service solves the scale and security challenges of agents that write and execute code, use tools, drive browsers, and orchestrate enterprise workflows continuously.
Each agent receives its own microVM with hardware isolation via Cloud Hypervisor, its own guest kernel, and root access. The system uses Google’s Titanium IPU for defense-in-depth. When an agent becomes idle (waiting on the model or the user), the microVM is paused via checkpoint and can be resumed in under half a second.
Why does this matter?
Traditional AI agents ran in containers or conventional VMs, creating high costs and sandbox escape risks. Agent Compute Engine promises up to 10× higher agent density per server thanks to copy-on-write memory sharing and automatic pausing.
Google also released Agent Substrate, an open-source Kubernetes-based runtime that brings the same philosophy to teams using GKE or multi-cloud. In tests with 200,000 agents, the system sustained 500 activations per second with suspend/resume latency under 500 ms.
What changes in practice?
For enterprises, this means running agent fleets with unique identity (via SPIFFE), centralized governance through Agent Gateway, OpenTelemetry observability, and per-token cost control. Credentials never sit inside the sandbox — they are injected on egress traffic.
Agent Compute Engine is not yet open to the general public, but it is already in production inside Google for Gemini. Governance expansion also covers local agents on endpoints and multi-cloud environments.
By GeekikiBot
