Vertex AI and Gemini Enterprise Agent Platform: the Map

GCP Basics · AI on GCP – Gemini, RAG, BigQuery ML and Agents · Lesson 1 of 5

Vertex AI, now delivered as Gemini Enterprise Agent Platform, groups models (Model Garden), tools to build and test prompts and agents, runtime for agents, and MLOps for custom models; most teams start by calling a hosted model.

Google’s AI services for developers grew up under the name Vertex AI. In April 2026 Google announced Gemini Enterprise Agent Platform as its evolution: Vertex AI capabilities are now delivered through the Agent Platform. APIs still live under aiplatform.googleapis.com, and many docs, roles and CLI commands still say “Vertex AI” or gcloud ai.

flowchart LR
  DEV["Your app or agent"] --> MG["Model Garden<br/>Gemini, Gemma, Claude, Llama"]
  DEV --> ST["Studio<br/>prompt design and testing"]
  DEV --> AG["Agent tools<br/>ADK, Agent Runtime, Memory Bank"]
  DEV --> RAG["Grounding and RAG<br/>search, vector search, RAG Engine"]
  ML["Custom ML models"] --> OPS["MLOps<br/>training, pipelines, registry, endpoints"]
  %% step DEV: Most projects start with an app or agent that calls a hosted model through an API.
  %% step MG: Model Garden lists 200+ models: Google’s Gemini and Gemma, partner models such as Claude, and open models.
  %% step ST: Studio is the console workspace to write prompts, compare models and export code.
  %% step AG: The Agent Development Kit and Agent Runtime build and host agents with sessions and long-term memory.
  %% step RAG: Grounding connects answers to Google Search or to your own documents and databases.
  %% step OPS: For your own trained models: training jobs, pipelines, Model Registry and online prediction endpoints.

What lives where

Area	Examples	Typical user
Models	Gemini (Pro, Flash, Flash-Lite), embeddings, image and speech models, Gemma, partner models	App developers
Build and test	Studio, prompt management, evaluation	Developers, product teams
Agents	Agent Development Kit (ADK), Agent Runtime, sessions, Memory Bank, Agent Garden templates	Agent builders
Grounding / search	Grounding with Google Search, Vertex AI Search, RAG Engine, Vector Search	Developers
MLOps	Workbench notebooks, custom training, Pipelines, Feature Store, Model Registry, endpoints, monitoring	ML engineers
Governance	Agent Identity, Agent Registry, Agent Gateway, Model Armor	Platform and security teams

Three ways to use AI on Google Cloud

Call a hosted model (Gemini through the API): no infrastructure, pay per token. Start here.
Use AI from data tools: BigQuery ML and AI functions run models from SQL (lesson 4).
Train or tune your own model: custom training on GPUs/TPUs and deployment to an endpoint, or self-hosting open models on GKE / Cloud Run with GPUs.

Responsible use basics

Google states that customer prompts and data sent to these models are not used to train Google’s models without permission.
Choose a location: the global endpoint gives the best availability; regional endpoints keep processing in a region for data-residency needs.
Every call is an IAM-checked API call and can be audit-logged.

gcloud services enable aiplatform.googleapis.com --project=dl-ai-dev
gcloud ai models list --region=asia-southeast1          # your models in Model Registry
gcloud ai endpoints list --region=asia-southeast1       # deployed endpoints

Command	Meaning
`gcloud ai custom-jobs list --region=…`	Custom training jobs
`gcloud ai endpoints describe ENDPOINT_ID --region=…`	Endpoint with deployed models

Traps: model names and versions change often, so read the current model list before hard-coding an ID; a self-hosted model on GPUs costs money even when idle, a hosted API does not; “Vertex AI” and “Agent Platform” refer to the same family of services.