Keep meaning and authority connected to the runtime.

An original candidate architecture

1. Experience and integrationExisting business channels → Apigee and authorised service adapters.
2. Context and meaningExisting repositories → Knowledge Catalog and selected Agent Platform context capabilities. Owners maintain definitions, provenance and permissions.
3. Reasoning and workflowAgent Platform models and Agent Runtime; Workflows for explicit service coordination where suitable.
4. Authority and effectConfigured IAM, appropriate workload or agent identity, human-review application and backend business checks.
5. Evidence and operationSelected platform lifecycle capabilities plus decision evidence, operational review and measured outcomes.

The layout is original. Arrows represent proposed integration work; no Google Cloud service is connected to this release.

Names changed; verify the intended capability

The current platform overview describes Gemini Enterprise Agent Platform as an evolution of Vertex AI. The earlier Agent Engine overview now redirects to its scaling and runtime documentation. Vertex AI Agent Builder documentation also remains available. Use the current API, region and feature documentation for an actual implementation; these labels are not treated as interchangeable products.

Knowledge Catalog, formerly Dataplex Universal Catalog, supports catalogue and governance work. This does not automatically produce the organisation’s ontology or validate its sources. IAM provides resource-access controls; the organisation still defines who may make each business decision.

Workflows coordinates configured services and APIs. Callbacks can pause and resume a workflow, but need a human-decision application and action-specific authority checks. Apigee provides API-management capabilities, rather than a business approval mechanism.

Controls and deployment choices still to make

• Confirm target-region support, exact release status, access conditions and quotas for every chosen feature.
• Select permitted models and data locations; test context freshness, retrieval and permission filtering separately.
• Define business vocabulary, ontology scope and accountable catalogue stewards.
• Implement authenticated review, expiry, rejection, safe resumption and exact-action checks outside generated reasoning.
• Cost model usage, runtime, storage, network transfer, integration, support and the people operating the service.
• Review recovery and operational ownership using the Google Cloud Well-Architected Framework.

No model, runtime or customer environment has been tested by this release. Pricing, preview eligibility and regional suitability remain deployment decisions. See the capability comparison; the columns show candidates, not verified equivalence.