Google is releasing Agent Compute Engine in preview on Google Cloud, a novel infrastructure originally built to power its Gemini AI agents. This service provides isolated microVM environments that combine container speed with VM-level security, designed to efficiently handle the rapid start-stop nature and security needs of autonomous AI agents at scale.
- MicroVMs provide strong isolation with fast startup and suspend/resume
- Persistent sandboxes enable root-level control for complex AI agent workflows
- Service targets scale, security, and efficiency for agentic compute workloads
Infrastructure signal
Google’s Agent Compute Engine introduces a hardware-isolated microVM to cloud AI infrastructure, blending container-like speed with virtualization-grade security. Each AI agent runs in a dedicated sandboxed environment that persists across sessions and supports root-level operations on a lightweight virtual machine.
This approach represents a shift from traditional VM or container usage, targeting the unique needs of agentic workloads that require rapid instantiation, suspension, and secure isolation. Inspired by internal systems that support Google’s extensive AI compute demands, the service is designed for large-scale deployment across diverse agent tasks like large-scale benchmarking and reinforcement learning.
Developer impact
For developers, Agent Compute Engine enables a new class of AI applications by offering persistent, secure compute environments with full kernel and root access. This allows agents to run complex code loops, manipulate filesystems, and utilize headless browsers, facilitating development of autonomous assistants, coding agents, and other workflows requiring deep system control.
The service’s ability to quickly pause and resume microVMs reduces compute waste and cost, while enabling session persistence that can last for weeks. This improves developer workflow by providing reliable, scalable environments to build, test, and run AI workloads without sacrificing isolation or flexibility, addressing gaps left by pure container or VM solutions.
What teams should watch
Teams focused on AI agent development, platform engineering, and cloud cost optimization should prioritize evaluating Agent Compute Engine. Its strong isolation and persistent microVMs offer opportunities to secure sensitive workloads while increasing efficiency by reducing startup overhead and idle costs.
Additionally, teams responsible for observability and infrastructure reliability will find value in this service’s design that supports rapid recovery and debugging of agent failures within isolated sandboxes. Platform teams managing large-scale AI experiments or continuous evaluation workflows should also monitor Google’s further rollout and integration of this compute model into their broader cloud ecosystem.