Google has introduced a limited preview of Agent Gateway for cross-cloud, extending its Gemini Enterprise Agent Platform governance tools to AI agents running outside Google Cloud. The November 11 release covers compatible Kubernetes environments, including Amazon EKS, Azure AKS, Red Hat OpenShift, and bare-metal clusters.
The central idea is to separate where an agent runs from how its activity is governed. Customers can leave agents in their existing environments while registering agents and Model Context Protocol servers and routing model, tool, and agent-to-agent traffic through local gateway proxies.
A management layer, not a new AI model
Google’s release notes describe access-policy enforcement, Model Armor safety policies, and logs and traces sent to Cloud Logging and Cloud Trace. This is infrastructure for controlling and observing agents, not the announcement of a new language model or a claim that agents become reliable simply by using a gateway.
The overview documentation describes three parts: Google-hosted management services, a cross-cloud control layer, and proxies operated in the customer’s clusters. Central policy definition is paired with enforcement in the local environment.
That division may appeal to organizations testing agents across several infrastructure providers. The potential benefit is a more consistent place to describe policies and inspect activity. Whether it simplifies an individual deployment depends on the integration work, existing controls, and operational requirements.
Customers still own important setup work
The setup guide requires a CNCF-conformant Kubernetes cluster, an existing service mesh or network stack, outbound connectivity to Google Cloud APIs, and suitable project permissions. It also calls for enabling supporting APIs and registering the external cluster.
These requirements make the preview a platform-engineering project rather than a switch that a business user can turn on. Security and infrastructure teams should review the identity arrangement, network paths, and operational ownership before placing sensitive workloads behind it.
Google’s shared-responsibility description leaves customers responsible for their external cluster and network configuration, including routing traffic into the proxies. Installing a gateway is not the same thing as proving that every relevant request actually passes through it.
Preview limits and costs need attention
Access requires a request, and the feature is not generally available. Google says the gateway itself carries no charge during limited preview, but associated services such as Model Armor, Cloud Logging, and Cloud Trace retain their standard charges.
The documentation also lists known issues, including a gap between a gateway reporting readiness and accepting connections, and restrictions involving service URL uniqueness across registered clusters. These details matter when designing tests and deciding whether a workload is appropriate for an early-stage service.
For a pilot, a useful first exercise would be to verify an allowed request, a blocked request, and an attempted route around the gateway. Teams should also test lost connectivity and confirm that logs contain the evidence their reviewers need. Those are practical evaluation steps, not outcomes Google has demonstrated for every customer.
The development is worth tracking because it addresses governance without requiring runtime migration. The next milestones are broader availability, operating experience, and clearer evidence about the effort required to manage agents consistently across infrastructure boundaries.



