Module 3 is a map. It divides the gen AI world into five layers, then walks down them one at a time, naming the Google Cloud product that lives at each. The reason to learn the layering rather than just the product names is that the exam’s favorite question shape is a scenario plus “which layer is this?” — and the layers are also how the course assigns roles, costs, and decisions. Knowing where something sits tells you who owns it and what it costs.1
The five layers
Top to bottom: Applications → Agents → Platform → Models → Infrastructure. Each layer consumes the one below it and hides that complexity from the one above.
- Gen AI applications — the user-facing frontend, where people actually interact with AI capabilities. The Gemini app, Google Workspace with Gemini, Gemini Notebook.
- Agents — software that learns how best to achieve a goal given its inputs and the tools available to it. The defining word is autonomous action: an agent does things, it doesn’t just answer. Customer agents, code agents, data agents.
- Platform — the tools and services that let agents and models interact: APIs, data management, model deployment. This is the layer that makes the others usable together. Agent Platform.
- Models — the brain. A complex algorithm trained on vast data that generates content, translates, and answers questions. LLMs, image recognition models, recommendation systems.
- Infrastructure — the core computing resources: the hardware and software needed to store and run models and training data.
Layer-identification drill
The exam gives a scenario and wants the layer. The tell is usually a single verb — training tools means Platform, generates instantly for a user means Application, processing power means Infrastructure.
| Scenario | Layer |
|---|---|
| Tools to label training data, experiment with architectures, deploy a finalized model | Platform |
| A website that generates an image instantly from a text description | Application |
| Needing more processing power and high-speed data connections to train a large model | Infrastructure |
| A system trained on millions of lines of code that suggests the next line | Model |
| Monitoring social media, identifying negative sentiment, and drafting a response automatically | Agent |
| A data science team needing powerful hardware for compute-intensive training | Infrastructure |
| Defining the behaviors and personalities of game NPCs that adapt to players | Agents |
The agents layer
A gen AI agent is an application that tries to achieve a goal by observing the world and acting upon it using the tools at its disposal. Module 5 covers how one is built internally; here the concern is what agents do and how they relate to the applications that contain them.
Three capabilities:
- Understand and respond to natural language — more intuitive interfaces that handle complex requests.
- Automate complex tasks — multi-step processes carried out within an application.
- Personalization — learn user preferences and tailor the experience.
Don’t confuse these: agents vs. applications
This distinction is tested repeatedly and in both directions. The agents layer defines the AI’s capabilities — searching, booking, summarizing, recommending. The applications layer provides the user-facing tool and the overall goal — the website, the chat window, the progress tracking. An agent is an intelligent component functioning inside a larger gen AI-powered application.
A related pairing, from the creative-writing-assistant activity: the model layer generates grammatically correct text; the agent layer identifies and applies the appropriate guidelines based on content type. Raw generation is the model; judgment about which rules apply is the agent.
Two agent categories
Both follow the same internal sequence — understand, call a tool, generate, deliver. What differs is the trigger and the output.
- Conversational agent — you provide input by typing or speaking; the agent understands, calls a tool, generates a response, and delivers it. Answering questions, casual chat, accessing information.
- Workflow agent — you define a task or trigger a process; the agent understands, calls a tool, generates a result, and delivers it. Ecommerce order fulfillment, customer onboarding, automated research, security log parsing.
Exam note: guiding new customers through account setup, tutorials, and FAQs is customer onboarding — a workflow agent, despite involving conversation. Match on whether a process is being executed, not on whether the interface is chat-like.
Agent use cases
Six categories the exam likes to match against scenarios: customer service (answer questions, resolve issues, recommend), employee productivity (find information, manage tasks, automate workflows), creative (generate ideas, create content, translate), code (write, review, debug, generate), data (analyze datasets, identify trends, extract insights), and security (automate security tasks).
Multi-agent systems
An application composes several specialized agents while supplying the shared interface. A travel booking app: one agent finds flights and hotels, another suggests activities, and the application provides browsing and reservation. The same shape appears in the customer support example (an agent answers, troubleshoots, and escalates to humans) and the personalized learning example (an agent assesses knowledge, recommends materials, generates exercises).
Reasoning frameworks
The reasoning loop often uses advanced prompt engineering frameworks to guide its decisions — simple rule-based calculations, complex thought chains, ML algorithms, or probabilistic reasoning. The two named patterns are ReAct (reasoning + acting) and chain-of-thought (CoT) prompting, both covered in full in Module 5.
Exam note: the two elements distinguishing AI agents from standalone LLMs are a reasoning loop and the ability to use tools. This is the single most repeated fact in the module — it appears in the lesson quiz and again in the wrap-up.
The platform layer
The platform layer is the foundation for building and scaling AI initiatives. Agent Platform streamlines the entire ML workflow by supplying infrastructure, pre-trained models, and tools to build, deploy, and manage. Its selling points: open and flexible, powerful infrastructure, pre-trained models, comprehensive tooling, customization, and easy integration.
MLOps tools — know which tool solves which job, because the exam tests exactly that discrimination:
| Tool | Job |
|---|---|
| Feature Store | Share, serve, and reuse ML features for consistency across models |
| Model Registry | Manage versions, track changes, organize models across their lifecycle |
| Model Evaluation | Evaluate and compare model performance |
| Workflow orchestration (Pipelines) | Automate the end-to-end ML pipeline, preprocessing through deployment |
| Model Monitoring | Detect performance degradation and input skew/drift, trigger retraining |
Exam note: the three scenarios and their answers — tracking experiments and model versions is Model Registry; a deployed model whose performance is declining is Model Monitoring; automating the whole pipeline is workflow orchestration. Registry records, Monitoring watches, orchestration runs.
The models layer
At the heart of every AI system is the model — a sophisticated mathematical structure trained on massive data. Agent Platform acts as a model hub offering two paths, and choosing between them is a recurring exam question.
Use existing models with Model Garden — discover, customize, and deploy from 160+ models, some usable out of the box. The right choice when you need something readily available and high-performing and don’t need control over architecture.
- First-party foundation models: Gemini, Nano Banana, Veo, Chirp.
- First-party pre-trained APIs: Speech-to-Text, Natural Language Processing, Translation, Vision.
- Open models: Gemma 2, CodeGemma, PaliGemma, Llama 3.1/3.2, Mistral AI, AI21, TII.
- Third-party models: Anthropic’s Claude family.
Build models with Agent Platform — either fully custom (create and train at scale, complete control over architecture and training) or AutoML, which creates and trains models with little or no code.
AutoML objectives by data type:
| Data type | Supported objectives |
|---|---|
| Image | Classification, object detection |
| Video | Action recognition, classification, object tracking |
| Tabular | Classification or regression, forecasting |
The standard model workflow is the same five stages as the ML lifecycle in Module 2: gather data → prepare data → train → deploy and predict → manage.
Exam note: a global ecommerce platform needing a readily available, high-performing translation model on a moderate budget → Model Garden. A researcher building a cutting-edge protein-folding model who needs complete control over architecture and training → build a custom model on Agent Platform. The discriminator is control, not sophistication.
The infrastructure layer
Infrastructure is the hardware and software providing the resources to train, deploy, and scale models. Three components:
High-performance computing
- GPUs and TPUs — the workhorses of AI, both excelling at parallel processing. GPUs provide general-purpose parallel processing power; TPUs are custom-designed and optimized specifically for AI tasks.
- Hypercomputers — supercomputers built by connecting many individual nodes, providing the massive scale needed to train and run gen AI models.
High-performance storage — Google Cloud’s storage is optimized for AI workloads with high throughput, scalability, and the ability to create dense compute clusters for faster training and inference. Fast read/write speeds matter because storage that can’t keep up starves the processors.
Networking — coordinating many processors requires fast, efficient communication; Google’s global fiber network supplies high-bandwidth, low-latency connectivity.
Exam note: GPUs and TPUs are specialized processors designed for parallel processing in AI tasks. High-performance storage exists to store and efficiently access the massive datasets used in training. And if you’re not on a managed platform, infrastructure must be considered at every development stage — data collection, training, deployment, monitoring, and refinement, all of them.
The travel chatbot, mapped through all five layers
The course’s integrating example: the application is the website and chat window; the agents are a conversational agent as the front door, a planning agent for flight search, and a booking agent; the platform is Agent Platform; the model is Gemini Pro; the infrastructure is the underlying compute, storage, and networking.
Edge computing
Edge computing runs AI on devices or servers closer to the data source or the point of need, rather than in a centralized cloud. The tradeoff it buys: you give up cloud scale to gain latency, privacy, and independence from connectivity.
- Real-time responsiveness — no network round trip.
- Increased data privacy — data stays on the device.
- Reduced reliance on internet connectivity — it works offline.
LiteRT (Lite Runtime) helps ML models run efficiently on edge devices and mobile phones. Gemini Nano is Google’s edge-designed model, delivering privacy, speed, and offline access. It powers Pixel’s Call Notes (summarizing phone conversations) and Pixel Recorder (summarizing voice recordings), and is available to Android developers through the AI Edge SDK.
Edge deployment on Agent Platform is three steps: convert models to LiteRT for optimal edge performance, package models and dependencies into containers for edge hardware, then manage and monitor deployments and gather insights.
Edge or cloud? Pick edge when latency is genuinely critical; pick cloud when you need scale or centralized aggregate analysis.
| Scenario | Answer | Why |
|---|---|---|
| Medical device analyzing patient data in real time during surgery | Edge | Cloud latency is unacceptable |
| Customer service chatbot handling millions of daily inquiries | Cloud | Needs scalable, robust infrastructure |
| Smart-city traffic pattern analysis and optimization | Cloud | Needs a centralized system with cloud-scale capacity |
Exam note: edge computing’s primary advantage is real-time responsiveness and reduced latency. Privacy and offline access are real benefits but are not the answer to “primary advantage.”
Roles mapped to layers
The course assigns each role a home layer, which is also how it frames who should be handed which product.
| Role | Responsibility | Layer |
|---|---|---|
| Business leaders | Interact with pre-built gen AI solutions to enhance operations and customer experience (Workspace with Gemini) | Applications |
| Developers | Build and deploy custom AI agents, integrate AI into existing applications (AI Applications, code generation, pre-trained APIs) | Agents |
| AI practitioners | Customize, deploy, and optimize models; scale workloads; implement responsible AI | Platform / Models / Infrastructure |
Exam note: AI practitioners’ primary responsibility is customizing, deploying, and optimizing generative AI models while ensuring responsible AI practices — the responsible-AI clause is part of the correct answer. Developers are the ones building and deploying custom agents.
Cost
Three paid activities: training the model, deploying it to an endpoint, and using it to make predictions. Compute time, storage for training data, and model outputs all factor in.
Pricing models:
- Usage-based — pay for what you use, measured in tokens or characters.
- Subscription-based — a recurring fee for access, with usage limits or feature tiers.
- Licensing fees — one-time or recurring, for commercial use or embedding in products.
- Free tiers — limited free access for experimentation or non-commercial use.
Pricing metrics: tokens (a piece of text — a word or part of a word), characters, requests (charged regardless of complexity), and compute time (significant for resource-intensive tasks).
Cost drivers: model size and complexity (larger and more capable costs more), context window (a larger window increases cost), specialized features (separately priced), and deployment (compute-based costs). Google Cloud’s pricing calculator is the tool the course points at.
Exam note: the primary cost associated with building gen AI solutions is model training. And when asked which factor increases the cost of using a model, the answer is a larger context window.
Time scales with customization: the more custom the solution, the more time and resources it takes to build.
Choosing a solution
Five dimensions to work through, plus a set of secondary constraints that often turn out to be the deciding ones.
- Scale — small scale means selecting pre-built tools and leveraging existing applications; large scale means designing for scalability and security while weighing infrastructure cost, data storage, and latency.
- Customization — start with existing models, then ask what’s genuinely unique: specialized knowledge, complex tasks, or a distinctive user experience. Fine-tune on domain-specific data only when the domain actually demands it.
- User interaction — UI (a dedicated interface, or embedded in existing workflows) and UX (conversational, informative, or task-oriented; how much guidance and feedback users need).
- Privacy — data security through encryption, access controls, and secure data centers; plus compliance with whatever regulations apply.
- Other considerations — latency tolerance, connectivity and offline needs, acceptable accuracy and error tolerance, and explainability (whether you need to understand the AI’s reasoning, which is non-negotiable in some domains).
Exam note: the key factor in deciding how custom a solution needs to be is whether grounding or fine-tuning is sufficient, or a brand-new model is required.
Solution-fit patterns from the activities: realistic images from text, some coding skill, short on time → a pre-trained API like Nano Banana. A recommendation chatbot with no coding experience → AI Applications to create a conversational agent. Product descriptions needed quickly by a marketing team → a gen AI-powered application like Google Workspace with Gemini. The pattern: the less time and skill available, the higher up the stack the answer sits.
Decisions and maintenance
Making decisions — evaluate model capabilities, compare pricing structures, factor in additional costs, and read the fine print. Sources for comparison: provider websites, research papers and benchmarks, and community forums.
Exam note: an important step in deciding on a gen AI solution is comparing different companies and models. The course treats single-vendor evaluation as a mistake.
Maintenance — five ongoing obligations: model monitoring and periodic retraining; regular data updates to keep the model fresh and relevant; software updates and bug fixes; hardware and infrastructure upkeep (server maintenance, security updates, capacity planning); and continuing security and compliance review.