Scale AI sets out Google Cloud architecture for enterprise agents

On September 22, Scale AI published a reference architecture that it described as a joint design with Google Cloud. The design covers deploying Scale GenAI Portfolio agents on Google Cloud and making them available through Gemini Enterprise. It gives technical teams a documented route from agent development to employee access while keeping the deployed workload in the customer’s cloud project.

In the proposed flow, teams build and evaluate an agent in Scale’s GenAI Portfolio, then deploy it inside the customer’s Google Cloud project using the customer’s virtual private cloud and encryption keys. The same underlying agent can serve a dedicated business application and appear in Gemini Enterprise without being rebuilt for each interface.

According to Scale, Google Cloud supplies access to Google and third-party models, identity and governance services, and infrastructure including Google Kubernetes Engine, GPUs and TPUs. Scale’s platform provides agent development, orchestration, deployment, evaluation, tracing and human-review capabilities.

The architecture uses the Agent2Agent protocol, or A2A, and the Model Context Protocol, or MCP, as interoperability interfaces, while Google’s Agent Registry provides discovery. A2A gives agents built on different platforms a common way to interact. Google Cloud’s A2A registration documentation confirms that administrators can connect externally hosted A2A agents to Gemini Enterprise and make them available in its web app. Google also documents how administrators can import Agent Registry agents into a Gemini Enterprise app for end-user access.

Scale said organizations can apply identity, data-loss-prevention, audit and spending controls through a shared Google Cloud governance framework. Google Cloud separately documents resource-level IAM controls for Gemini Enterprise apps and data stores: a user needs permission for both the app and the target data store to receive answers from that data store. Its documentation also identifies Cloud Audit Logs for administrative and data-access activity.

The controls depend on the integration path. Google says A2A agents registered directly with Gemini Enterprise do not send traffic through Agent Gateway, so gateway policies do not apply to that route. For agents imported from Agent Registry, governance policies apply only to agents in the registry associated with the configured gateway. Direct communication between Gemini Enterprise agents, or between an agent and an MCP-based data connector, does not trigger gateway policy enforcement.

Scale did not name any customer deployments or provide pricing, benchmark results or measured production outcomes for the architecture. Its post also did not link to a separate downloadable architecture document beyond the high-level diagram in the announcement.