Prepare and test your skills
Prepare and test your skills
Vertex AI Agent Builder and Orchestration is the platform layer that turns chosen models into deployable AI agents and manages those agents through deployment and operation. In Google Cloud, this work is handled by the Gemini Enterprise Agent Platform family, which provides builder interfaces, orchestration frameworks, and serving and governance services. Grounding connects those agents to enterprise and real-world data so answers stay factually accurate and relevant to the business.
Building applications on the platform means combining low-level API calls into a high-level workflow that can be deployed as one unit. Gemini Enterprise Agent Platform Vision (opens in a new tab) shows this pattern for video and image solutions, letting users build and deploy applications through a simplified user interface. The Vision builder depends on three connected parts, with Live Video Analytics and data streams supplying input, analysis models producing insights, and Vision Warehouse supplying searchable storage.
The platform supports three core builder capabilities:
The Vision workflow runs those capabilities in a fixed order:
A three-step process flow showing real-time video ingest, orchestration to general custom and specialized models, then storage in Vision Warehouse or BigQuery or live receiver, deployed by one Vision API request.
The Vision API creates and updates a high-level workflow that combines multiple individual API calls, then executes the workflow as a unit when the team makes a single deploy request to the platform server.
Orchestration on the platform coordinates models, features, pipelines, and predictions across development and production. Gemini Enterprise Agent Platform Pipelines automates, monitors, and governs machine learning systems, while Gemini Enterprise Agent Platform Inference serves predictions from trained models. The lifecycle also depends on supporting services where Feature Store streamlines feature management and online serving, Model Registry manages the complete lifecycle of models, Model Monitoring watches for feature skew and drift, and Vertex AI TensorBoard tracks and compares experiments.
Cloud Run (opens in a new tab) and Google Kubernetes Engine (GKE) provide two infrastructure choices for running AI and machine learning workloads:
| Option | Role in orchestration | Choose when |
|---|---|---|
| AI/ML orchestration on Cloud Run | Runs AI/ML workloads directly on scalable infrastructure | The need is to run workloads without managing Kubernetes clusters |
| AI/ML orchestration on GKE | Orchestrates the complete AI/ML lifecycle with Kubernetes | The need is to orchestrate complex, container-based training and serving with Kubernetes control |
| Gemini Enterprise Agent Platform Pipelines | Automates, monitors, and governs ML systems | The need is repeatable, governed pipelines rather than one-off runs |
Multi-agent systems add a network boundary to orchestration. The app itself can be publicly accessible while private connections carry traffic between agents, subagents, and tools, so public users sit on one side of the boundary and internal agent-to-tool calls sit on the other. This pattern supports uses such as orchestrating investigation and triage in a security operations center and providing technical guidance through continuous, bidirectional multimodal streaming.
Exam tip: Pipelines automates and governs the system, while Inference serves predictions from an already trained model.
Grounding is the practice of connecting an agent's responses to trusted outside data instead of relying only on what the model learned during training. A grounded agent pulls context from enterprise sources and real-world services at query time, so factual accuracy improves and answers reflect current business data. In Google Cloud, grounding data flows from raw silos through a highly governed, secure pipeline into AI and agentic actions, which creates a borderless open data lakehouse from ingestion to action.
Enterprise grounding commonly uses retrieval-augmented generation with a vector store. One architecture runs retrieval-augmented generation using GKE with Cloud SQL, while another uses AlloyDB for PostgreSQL as the vector store for a generative AI application. Multimodal GraphRAG extends the same idea by consolidating fragmented multimodal data into a searchable knowledge graph, so the agent retrieves connected facts rather than isolated passages. Secure data environments protect this flow by shielding sensitive datasets against accidental exposure and malicious exfiltration.
Trusted agents also ground in real-world contextual data from Google Maps Platform and Google Calendar. Data flows from Maps Platform and Calendar into the agent, so location, place, and scheduling context shapes the response. Classification systems use the same principle when a multi-agent system analyzes disparate multimodal data and produces a high-confidence classification, because the final decision depends on grounded inputs from many sources rather than a single model guess.
Generative AI Studio (opens in a new tab) is a Google Cloud tool for designing, testing, and customizing prompts sent to Gemini and PaLM 2 large language models (LLMs). Generative AI Studio sits inside Vertex AI, which provides access to Google's large generative AI models so users can test, tune, and deploy them. Generative AI Studio is one of the gen AI tools available in the Cloud Console, alongside Colab Enterprise, Notebooks, Workbench, and SDKs and APIs.
Prompt design is the process of authoring prompt and response pairs to give language models added context and instructions. In Generative AI Studio, users design, test, and customize prompts for Gemini and PaLM 2 models. After prompts are authored, they can be fed to the model as a prompt dataset, and when the model serves predictions it responds with those instructions built in. Prompt engineering in the development flow includes the prompt gallery, Vertex AI Studio, compare prompts, and optimize prompt, so users can compare and refine prompts before tuning or deployment.
Tuning is the stage where a generative model is adjusted to perform specialized tasks better. If a use case is specialized, model performance can be improved by using model tuning such as supervised tuning for classification, sentiment analysis, or entity extraction. Models can also be tuned further with data examples that show the desired behavior. Optimization includes distillation, which is a separate stage that refines efficiency, and together prompt engineering, tuning, and optimization form a looping evaluation subcycle with model selection.
Exam tip: Use prompt engineering in Generative AI Studio to design and test prompts; use tuning when a specialized task needs improvement with data examples.
The generative AI development flow is the ordered lifecycle that moves work from idea to production. The flow contains six stages, where the first four repeat inside an evaluation loop until results are good enough:
Generative AI Studio is most directly used in the prompt engineering stage, but its outputs feed into tuning and deployment. A personalized marketing prototype shows how the stages connect to Google Cloud services: user data flows from Google Cloud and external sources into BigQuery, a Dataflow pipeline derives marketing insights, Eventarc triggers a Cloud Run service, the Cloud Run service sends those insights with a prompt to the Gemini Enterprise Agent Platform, the Gemini model generates audio, video, and text content, and Cloud Run uploads the content to a Cloud Storage bucket. When users visit the company web portal hosted on Cloud Run, the portal retrieves user-specific content from Cloud Storage and displays it, with an optional feedback loop and human review step before upload to improve quality.
Choosing between generative AI and traditional AI depends on whether the goal is to create new content or to predict outcomes from historical data. Generative AI is suited to creating and recommending content (opens in a new tab), powering conversational search and chatbots, scaling repetitive workflows, finding insights in documents, and generating code. Traditional AI is better for classification, regression, time series forecasting, and clustering based on historical data.
| Approach | Best for | Examples | Data needs |
|---|---|---|---|
| Generative AI | Creating new content, summaries, conversations, and multimodal outputs | Summarization, content generation, advanced transcription, chatbots | Can use prompt engineering when training data is unavailable or too small; can be tuned with data examples |
| Traditional AI | Predicting outcomes or classifying categories from historical data | Spam filtering, defect detection, house price prediction, sales forecasting, customer segmentation | Often uses structured data and historical datasets; may need custom training for more control |
If a classification or detection use case can be met by a pre-trained traditional AI model, use that model. If not and enough training data exists, choose custom training when high control over training is needed, or choose generative AI when faster go-to-market is the priority. If training data is unavailable or too small, use generative AI models with prompt engineering and tune them with data examples. The two approaches can also combine, where output from a traditional AI model becomes part of the prompt for a generative AI model.
Vertex AI Model Garden (opens in a new tab) is an AI and machine learning model library inside Vertex AI that brings Google models, partner models, and curated open-source models into one place. Model Garden and Discovery is the starting point for generative AI work, where teams find and compare foundation models before building applications. Discovery flows from Model Garden to testing, tuning, deployment, and monitoring, so model selection connects directly to application deployment.
Model Garden is a central library that helps teams discover, test, customize, and deploy models and assets from Google and Google partners. All available models are grouped in a single location with a consistent deployment pattern, so different model types follow the same basic path to an endpoint. Model Garden provides built-in integration with tuning, evaluation, and serving, and the platform handles deployment and serving of generative AI models that would otherwise be difficult to operate.
Discovery in Model Garden happens on the Model Garden page in the Google Cloud console, where teams browse foundation, tunable, and task-specific models. The filter pane narrows the catalog by business need and source before a team opens a model card for details:
After filtering, each model card describes what the model does and how it can be tested, customized, and deployed. This discover-then-evaluate pattern keeps selection tied to real requirements instead of choosing a model by name alone.
Model types in Model Garden differ by readiness and customization path. The catalog spans first-party Google models such as Gemini and Imagen, featured third-party partner models, and select open-source assets including Hugging Face text-generation models. Google Cloud provides state-of-the-art foundation models through Vertex AI, including Gemini, while Hugging Face text-generation models can be deployed to Vertex AI or self-hosted.
| Category | What it offers | How you customize it |
|---|---|---|
| Foundation models | Pretrained multitask large models | Tune or customize for specific tasks using Agent Studio, Agent Platform API, and Agent Platform SDK |
| Fine-tunable models | Models prepared for additional training | Fine-tune using a custom notebook or pipeline |
| Task-specific solutions | Prebuilt models for defined tasks | Use ready to run, with many models customizable using your own data |
The generative AI development flow in Model Garden moves work through an ordered lifecycle from idea to production. Data and decisions flow from selection forward to deployment, with early stages repeating until results are good enough:
The first four stages form a looping evaluation subcycle, so selection, prompting, tuning, and optimization repeat before deployment. For hosting, a third-party model can be deployed to Vertex AI Model Garden or self-hosted on Google Kubernetes Engine (GKE) or Compute Engine, and Compute Engine virtual machines can attach GPUs to accelerate generative AI workloads. Safety surrounds this lifecycle: Google performs testing and benchmarking on the serving and tuning containers it provides with active vulnerability scanning of container artifacts, third-party models from featured partners receive checkpoint scans for authenticity, and Hugging Face Hub models are scanned for malware, pickle files, Keras Lambda layers, and secrets. Access is controlled with a Model Garden organization policy that applies at the organization, folder, or project level, while cost follows compute use because open-source tuning is charged like custom training and deployment to an endpoint is charged for prediction compute.
Exam tip: Hugging Face models deemed unsafe are blocked from deployment in Model Garden, while models deemed suspicious or able to potentially execute remote code are still deployable but flagged for thorough review.
Gemini Enterprise Agent Platform Pipelines automates, monitors and governs machine learning systems, while Gemini Enterprise Agent Platform Inference serves predictions from trained models. Pipelines is used for repeatable, governed pipelines rather than one-off runs.
Use AI/ML orchestration on Cloud Run when the need is to run AI and machine learning workloads without managing Kubernetes clusters. Use AI/ML orchestration on Google Kubernetes Engine (GKE) when the need is to orchestrate complex, container-based training and serving with Kubernetes control.
Grounding is the practice of connecting an agent's responses to trusted outside data instead of relying only on what the model learned during training, so factual accuracy improves and answers reflect current business data. Enterprise grounding commonly uses retrieval-augmented generation with a vector store such as Cloud SQL or AlloyDB for PostgreSQL and can extend to Multimodal GraphRAG knowledge graphs, plus real-world context from Google Maps Platform and Google Calendar.
Generative AI is suited to creating new content, summaries, conversations and multimodal outputs such as content generation, chatbots and code generation, and it can use prompt engineering when training data is unavailable or too small. Traditional AI is better for predicting outcomes or classifying categories from historical data such as spam filtering, defect detection, house price prediction, sales forecasting and customer segmentation.
An enterprise development team is designing a multi-model generative AI application on Google Cloud. The team needs a centralized environment to discover, compare, and test various foundation models—ranging from proprietary Google models to third-party offerings and curated open-source models—via managed APIs and dedicated model cards before selecting one for production.
Which Google Cloud capability should the team use?
Cloud Storage
Vertex AI Feature Store
BigQuery ML
Model Garden
Last updated