This is the platform layer, and the first thing to fix is the name.
The rename
Vertex AI is now Gemini Enterprise Agent Platform, usually shortened to Agent Platform. Vertex AI Search is now Agent Search. Google updated the exam for the new naming, and a great deal of study material still uses the old one. Learn both, answer with the new.
Model Garden
The catalogue. Google's own models, open models such as Gemma, and third party models, all reachable through one platform with one set of controls.
The business value is comparison without commitment. A team can evaluate several models against its own task and switch later without rebuilding everything around the call. It is also the concrete thing behind Google's open approach claim, since third party and open models sit beside Google's own.
Agent Platform AutoML
Training a custom predictive model from your own labelled data without choosing an architecture or writing training code. You supply examples, it produces a model.
This is classical machine learning rather than generative, and the exam includes it because the right answer to a business problem is often a prediction rather than a generation. Forecasting demand, scoring leads, classifying records. A question describing labelled historical data and a number or category to predict is pointing here, not at Gemini.
The RAG offerings
Two, and the difference is how much you assemble.
Prebuilt RAG with Agent Search is the managed path. Point it at your content and it handles chunking, embedding, indexing and retrieval. Configuration rather than construction, and the right default.
RAG APIs expose the stages so you can control them. Choose your own chunking strategy, your own embedding model, your own vector store, your own ranking. Worth it when retrieval quality is the product and the managed defaults are not good enough, and not before.
The order matters. Teams routinely build a retrieval pipeline by hand and discover afterwards that the managed one would have done.
Custom agents
Agent Platform is also where agents get built. An agent is given a goal, a set of tools it may use, and instructions on how to behave, and it then plans and acts rather than answering once.
The business value is work that has steps. Not answering a question about an order, but finding the order, checking the policy, issuing the refund and writing to the customer.
The cost is that the security questions change. An agent that can act on systems means a prompt injection reaches something that changes data, so what the agent is permitted to do matters more than what it is told.
Agent Studio and Google AI Studio
Two build environments with different audiences.
Google AI Studio is the fast way to try the Gemini API. Prompt in a browser, get code out, prototype in an afternoon. It suits a developer testing an idea.
Agent Studio, within Agent Platform, is where an agent is designed, given its tools, tested and deployed under enterprise controls.
Prototyping points at Google AI Studio. Building something that will run in a business points at Agent Studio.