Gen AI Leader
Deep Dive
While the previous products are aimed at end users or business users, Agent Platform is the layer aimed at developers: access to models, enterprise search reusable as a component, and tools for building custom agents from scratch.
Contents
The Agent Platform (previously known as Vertex AI Agent Builder / Vertex AI Search) is Google Cloud's set of tools for developers to build, train, and deploy generative AI solutions, from access to models all the way to full agents in production. It is organized into three main pieces.
Model Garden
Model catalogDiscover, test, and deploy first-party, open, and third-party models from a single place.
Model Garden is the centralized catalog of foundation models available on Google Cloud: first-party models (Gemini, Imagen, Veo), open models (Gemma), and third-party models, all accessible through a unified discovery, testing, and deployment experience.
Why it matters
A developer doesn't have to search across multiple providers or manage separate credentials — Model Garden centralizes the discovery and evaluation of which model fits each task (text, image, video, embeddings) within the same Google Cloud project.
Agent Search
Reusable searchGoogle-grade search and retrieval capabilities, exposed as a component to incorporate into a custom agent.
We already saw Agent Search as a product oriented toward customer experience. From the Agent Platform perspective, it is also a component that a developer can incorporate directly into their own application or agent: Google-grade search and retrieval capabilities, exposed as a reusable service instead of having to build them from scratch.
Agent Platform AutoML
Custom MLTrain your own classification, regression, or vision models on your labeled data, with minimal code.
Agent Platform AutoML allows training custom machine learning models (classification, regression, vision) on the customer's own data, with minimal code: the developer provides the labeled data and AutoML handles the architecture selection and hyperparameter tuning.
When to use it instead of a foundation model
AutoML makes sense when the task is a classic, well-scoped ML problem (for example, classifying images of factory defects into predefined categories) with enough labeled data of your own, rather than an open-ended natural language task where a pretrained foundation model like Gemini already performs well without additional training.
RAG (Retrieval-Augmented Generation) lets a model ground its responses in real documents retrieved at query time, instead of relying solely on its parametric knowledge. Google Cloud offers two ways to implement it, with different levels of control.
Prebuilt RAG with Agent Search
A managed solution: the developer provides the documents (or connects them from an existing source) and Agent Search handles ingestion, chunking, indexing, and retrieval automatically. Less configuration, faster time to production.
RAG APIs
Granular control: the developer designs their own RAG pipeline using individual APIs (embeddings, Vertex AI Vector Search, custom chunking), choosing each piece based on specific performance or data-format needs.
How to decide between the two — a typical exam question
If the priority is implementation speed and minimal maintenance, use prebuilt RAG with Agent Search (managed). If the priority is fine-grained control over each stage of the pipeline (a specific chunking strategy, your own vector search engine, custom reranking logic), use the RAG APIs.
Beyond using models or search in isolation, Agent Platform lets you assemble full agents: systems that combine a base model, tools for acting on the environment (function calling, data connections), and orchestration logic to complete multi-step tasks, not just answer a single isolated question.
Example
An expense-approval agent built on Agent Platform can read a receipt (vision), check the company's expense policy (RAG over internal documents), verify the available budget (a call to an internal function/API), and decide to approve or escalate the request — all in a single automated interaction.
Did you get this topic?
Test what you just learned
A development team needs to implement RAG over an internal knowledge base, but requires granular control over the chunking strategy and the specific vector search engine used, because its document format is unconventional. Which Google Cloud option is best suited?
Put what you learned to the test with original Gen AI Leader exam questions.
Start Gen AI Leader mock examUnlimited mock exams with the Pro plan
Log in to track your progress.