Google Cloud strengthens secure hosting of AI agents
Google Cloud is preparing Agent Compute Engine, an infrastructure designed to run AI agents in isolated microVMs. The service aims to enhance security, reduce idle costs, and simplify access management.
Google Cloud has introduced Agent Compute Engine, an infrastructure designed to run artificial intelligence agents in an isolated environment. The service, announced for preview in the coming months, targets companies that want to operate agents capable of executing code, browsing in a browser, and orchestrating workflows, without turning their cloud into a permissions sieve.
Des microVMs pour isoler les agents sans les ralentir
At the core of the system, Google Cloud relies on microVMs, that is, lightweight virtual machines that maintain the hardware isolation of a hypervisor while being more efficient than traditional VMs. The company places them on stateless appliances, equipped with its Titan chips, to verify firmware integrity and limit the risk of compromise between environments.
The promise is simple: an agent can work in its sandbox, but it must not be able to escape to the rest of the system. Access to storage remains strictly limited, while network traffic passes through mediation layers and gateways that validate requests before any access to the data center. Even credentials are not left in the agent's environment: a dedicated gateway, Agent Gateway, transmits them on its behalf when an external service needs to be contacted.
Quand les agents dorment, la facture respire un peu
Google is not only selling security; the company is also looking to reduce the operating cost of these persistent agents. They often wait between tasks, but traditional sandboxes continue to consume resources during these times. Agent Compute Engine detects these idle periods, puts the environment into hibernation, then quickly reactivates it when needed.
Google Cloud engineers Ramanathan Venkatapathy and Rohit Jnagal explain that "This lets a single physical node host orders of magnitude more living agent microVMs than conventional VMs." The architecture must also allow snapshotting, cloning, and rollback without unnecessarily duplicating resources. In short, agents can sleep without monopolizing the server — a small budgetary miracle, provided the theory holds in production.
Un chantier qui va compter pour les entreprises
The arrival of this type of infrastructure shows that AI agents are no longer just laboratory demonstrations. As soon as they handle business applications, credentials, and workflows, the question is no longer just about the model, but also about the execution environment. Google Cloud responds with a stack designed for enterprise constraints: isolation, access control, flow supervision, and cost management.
The company also plans an enhanced version of Agent Gateway, with intelligent model routing and token management tools to monitor spending and apply budget caps. For IT and security teams, the topic is less glamorous than an agent that "does everything on its own," but much more concrete: it will determine whether these systems move from gimmick to truly deployable service at scale.
By the numbers
- Coming months — period announced for preview availability, according to Google Cloud.
- Orders of magnitude higher — hosting capacity mentioned by Google for a physical node.
- 2 proxy layers — local mediation and external gateway for network flows.
- 1 dedicated gateway — Agent Gateway manages credentials outside the sandbox.