TL;DR — Key Takeaways
- Google Cloud introduces infrastructure for AI agents: Agent Compute Engine uses hardware-isolated microVMs to support enterprise agents that execute code, operate browsers and manage workflows.
- Security controls extend beyond the sandbox: The architecture keeps long-lived credentials outside agent environments and enforces policies on network traffic and storage access.
- Idle sandboxes reduce resource consumption: Checkpointing preserves inactive environments so they can resume quickly, helping reduce the compute resources needed for persistent agents.
Google Cloud has unfurled Agent Compute Engine, a specialized infrastructure service designed specifically to run, secure, and scale enterprise artificial intelligence (AI) agents.
As businesses increasingly deploy AI agents to automate complex workflows, hosting these autonomous systems presents unique cybersecurity and resource challenges. AI agents typically operate inside secure digital environments known as sandboxes, which isolate applications, data, and user commands while enforcing cybersecurity guardrails.
Google Cloud’s new service powers these sandboxes using lightweight, isolated technology called microVMs. According to the company, microVMs combine the hardware efficiency of containers with the rigorous hardware-level isolation of traditional hypervisors, preventing malware from spreading between systems.
To deliver maximum isolation, Google hosts these microVMs on stateless, custom-built appliances equipped with Google’s proprietary Titan security chips. These chips continuously verify firmware integrity, while the underlying CPUs feature built-in mitigations against side-channel attacks, cyberthreats that attempt to steal data by analyzing a processor’s physical state.
Security controls extend throughout the host environment.
MicroVMs maintain strictly limited storage access, ensuring attackers cannot read or delete data even if a sandbox is compromised.
All network traffic passes through point-to-point IPC boundaries and dual proxy layers (a local mediation proxy and an external gateway) to strictly validate requests before reaching the broader data center.
Sensitive login credentials are stored outside the sandbox. When an agent requires access to an external app, a secondary service called Agent Gateway securely submits the credentials on its behalf.
Beyond security, Agent Compute Engine targets high operational costs associated with hosting autonomous software. AI agents frequently pause between tasks, yet traditional idle sandboxes continue consuming costly hardware resources. Google solves this by automatically detecting idle states and hibernating the sandboxes, re-instantiating them instantly when triggered.
“This lets a single physical node host orders of magnitude more living agent microVMs than conventional VMs,” Google Cloud Distinguished Engineers Ramanathan Venkatapathy and Rohit Jnagal wrote in a technical blog post. They said the architecture supports efficient snapshotting, cloning, and rollbacks without duplicating resource overhead.
Google plans to roll out Agent Compute Engine over the coming months alongside an upgraded version of Agent Gateway. The updated gateway will introduce intelligent model routing — dynamically directing tasks to the most cost-effective AI model — and new token management tools to help enterprises track spending and enforce budget controls.
