Why a map comes first
If you've used Google Cloud for a while, you already know the tricky part: the names keep changing. In April 2026, at Google Cloud Next '26, Google brought its AI platform together under one new name. Some blog posts, Terraform modules and Stack Overflow answers now use the new names, and many still use the old ones. Before you can choose a service, you need to recognise it under either name.
Think of this level as the station map on the wall of a metro system. It shows every line and every station, and which stations have been renamed. It doesn't tell you which route is best. That comes in later levels.
The rename you must know about
On 22 April 2026 Google introduced the Gemini Enterprise Agent Platform as "the evolution of Vertex AI". Most of Vertex AI's sub-products were renamed along with it. Google's official name-changes page lists over 50 of them. These are the ones you'll meet most often:
| Former name | Current name |
|---|---|
| Vertex AI (Platform) | Gemini Enterprise Agent Platform ("Agent Platform") |
| Vertex AI Studio | Agent Studio |
| Vertex AI Search | Agent Search |
| Vertex AI Agent Engine | Agent Runtime |
| Vertex AI Model Garden | Agent Platform Model Garden |
| Vertex AI Training | Agent Platform Managed Training |
| Vertex AI Inference / Prediction | Agent Platform Inference |
| Vertex AI Pipelines, Workbench, Model Registry, Vector Search, Feature Store | Same names, now "on Gemini Enterprise Agent Platform" |
| Google Agentspace (renamed October 2025) | Gemini Enterprise |
| Dialogflow CX | Conversational Agents (Dialogflow CX) |
Gotcha: don't mix up the two "Gemini Enterprise" names. Gemini Enterprise (formerly Agentspace) is a workplace app that staff use to search company data and use agents. The Gemini Enterprise Agent Platform (formerly Vertex AI) is the developer platform you build on. They share a brand but serve different people.
The good news for your code: this was a rebrand, not a migration. The API endpoint is still aiplatform.googleapis.com, and existing Vertex AI SDK calls keep working unchanged. Your Terraform google_vertex_ai_* resources keep working too. Nobody is forcing you to rewrite anything. New features arrive through new SDKs, which you can adopt when you're ready.
The map: six groups
1. Generative AI platforms and models
- Agent Platform (formerly Vertex AI): the managed platform for calling, tuning and serving models.
- Gemini models via the Agent Platform: text, code, images, audio and video in one model family.
- Model Garden: a catalogue of first-party, partner and open models.
- Agent Studio (formerly Vertex AI Studio): a console workspace for testing prompts and tuning.
Jobs: chat, summarising, code help, content generation, extracting data from documents with a prompt.
2. Agents
- Agent Runtime (formerly Agent Engine): managed hosting for agent code, with sessions and memory.
- Gemini Enterprise (formerly Agentspace): a ready-made agent and search app for employees.
- Conversational Agents (Dialogflow CX): chat and voice bots built around designed conversation flows.
Jobs: multi-step assistants, customer-service bots, internal helpdesks.
3. Search and retrieval
- Agent Search (formerly Vertex AI Search): managed enterprise search and grounding over your own content.
- RAG Engine and Vector Search: building blocks for your own retrieval pipeline.
Jobs: searching a website or intranet, giving answers that cite company documents.
4. Ready-made AI APIs
You call a pre-trained model, and you don't need any ML skills:
- Document AI: OCR and form or invoice parsing.
- Speech-to-Text and Text-to-Speech.
- Cloud Vision API: labels, OCR, detection in images.
- Cloud Translation.
- Cloud Natural Language: entities, sentiment, classification.
Jobs: perception tasks with clear inputs and outputs, priced per request or per unit of content.
5. ML platforms
- Agent Platform Managed Training, Inference, Pipelines, Model Registry, Workbench and Feature Store: the classic MLOps toolkit.
- BigQuery ML: train and run models with SQL, right where your data already lives.
Jobs: churn prediction, demand forecasting, fraud scoring, using your own tabular data.
6. Accelerators
- Cloud TPU: Google's own AI chips. The current generation is Ironwood (TPU7x).
- GPUs on Compute Engine and GKE.
Jobs: large training runs and high-volume self-hosted inference.
Walkthrough: see the map in your own project
API service names barely change, even when product names do. That makes them the most reliable way to find each service. You can list them like this:
# Which AI-related APIs are already enabled in this project?
gcloud services list --enabled \
--filter="config.name~(aiplatform|discoveryengine|documentai|speech|texttospeech|vision|translate|language|dialogflow|bigquery|tpu)" \
--format="value(config.name)"
Realistic output:
aiplatform.googleapis.com # Agent Platform (still the old endpoint name)
bigquery.googleapis.com # BigQuery, which includes BigQuery ML
discoveryengine.googleapis.com # Agent Search and Gemini Enterprise
documentai.googleapis.com # Document AI
Notice that none of these says "Agent Platform". A few groups map to their API like this:
# group -> API service name
# generative + ML plat. -> aiplatform.googleapis.com
# search / Gemini Ent. -> discoveryengine.googleapis.com
# conversational agents -> dialogflow.googleapis.com
# accelerators (TPU) -> tpu.googleapis.com
Best practice: in design documents, write the current name and the former name side by side, for example "Agent Search (formerly Vertex AI Search)". Add the API service name too. That way a reviewer who only knows one of the names can follow along.
What's retiring or closed
A map is no use if it sends you to a station that is about to close. Watch for these:
- Feature Store (Legacy): deprecated on 17 February 2026. Since 17 May 2026 it gets only critical patches, with no new features. It shuts down on 17 February 2027, and its APIs stop working then. Migrate to Feature Store (V2), which serves features from BigQuery.
- Feature Store Optimized Online Serving: the same shutdown date, 17 February 2027.
- Vertex Explainable AI: shuts down on 16 March 2027.
- Already gone: AutoML Text, legacy AI Platform Training, and Workbench managed and user-managed notebooks.
Trade-off: "Deprecated" still means it works today. If you start a new project on a deprecated service, the cost of migrating comes back within months. Check the deprecations page before you commit.
Key takeaways
- Six groups: generative platforms and models, agents, search and retrieval, ready-made AI APIs, ML platforms, accelerators.
- Vertex AI became the Gemini Enterprise Agent Platform on 22 April 2026. Studio became Agent Studio, Search became Agent Search, and Agent Engine became Agent Runtime.
- Agentspace became Gemini Enterprise in October 2025. That is the workplace app, which is a different product from the developer platform.
- Existing code keeps working because the endpoint is still
aiplatform.googleapis.com. - Feature Store (Legacy) shuts down on 17 February 2027, so plan for Feature Store (V2).