Prepare and test your skills
Prepare and test your skills
Agent tooling is how a generative AI agent moves beyond frozen training data to use live systems and current information. A language model alone cannot know today's news, current database rows, or live customer account state. Tools close that gap by letting the agent retrieve real-time data or perform an API-driven action, then use the returned result to finish the task.
The three core tool categories are extensions, function calling, and data stores, and each answers a different need. Extensions connect an agent to external systems and live APIs. Function calling lets an agent run a defined operation or custom code, such as a calculation or workflow step. Data stores ground answers in private organizational data.
| Category | What it connects the agent to | Example in Google Cloud |
|---|---|---|
| Extension | External systems and live APIs | Data Agent Kit extension for VS Code; Spanner MCP server; Gemini Enterprise for Customer Experience (opens in a new tab) tools |
| Function calling | Defined operations and custom code | Named tools in Gemini Enterprise for Customer Experience; n8n workflows on Cloud Run |
| Data store | Private organizational data | Conversational data agents (opens in a new tab) over Spanner |
Data flows from the external API or database into the model prompt, then from the model to the user as a grounded answer or action.
Exam tip: A data store supplies private data for grounding, an extension reaches a live external API, and function calling runs a defined action or custom code.
Grounding is the practice of connecting an agent to verifiable information before it answers, so outputs are based on facts instead of only training data. Retrieval-augmented generation (RAG) is the method that pairs grounding with generation. Data flows from an external data store or API into the model prompt, which improves accuracy and reduces hallucinations.
The RAG workflow runs in a fixed order. First, the user query is converted into an embedding and compared against indexed data. Then the system retrieves the most relevant documents, records, or API results. Only after retrieval does the model receive the original query plus the retrieved facts to generate the final answer.
A five-step left-to-right process flow showing a user query being embedded, searched, retrieved from indexed data, added to the prompt, and used to generate a grounded answer.
Google Workspace Google Search (GWGS) (opens in a new tab) grounds companion agent responses in real-time Google Search results. Grounded results are built from three inputs: the end-user prompt, customer-provided context, and Google Search engine data. This lets the agent answer accurately about current news and dynamic subjects. Enabling GWGS carries a tradeoff because Customer-Managed Encryption Keys (CMEK), Data Residency (DRZ), VPC Service Controls (VPC-SC), and Access Transparency (AXT) do not apply to the search requests.
Embeddings are the numerical foundation that lets retrieval find meaning rather than only matching keywords. An embedding converts text, images, or video into an array of floating point numbers called a vector, where closer proximity in vector space means more similar meaning. When a query arrives, the query embedding is compared to document embeddings through similarity search.
Many search engines store embeddings in a vector database, which is optimized for similarity search at scale. The vector database depends on the embedding model that created the vectors, because queries and documents must use compatible representations to match correctly. Google Cloud offers managed options for different RAG needs.
| Option | What it provides | Choose when |
|---|---|---|
| Vertex AI Search | Managed search engine over many data sources | A ready search engine is needed without managing embeddings |
| Vertex AI RAG Engine | Managed RAG orchestrator | A customizable retrieval pipeline is needed |
| Vertex AI Vector Search | Optimized vector database | Low-latency, large-scale similarity search is needed |
Agents can also ground where trusted data already lives. AlloyDB and Cloud SQL support working with embeddings for operational data that needs low latency, while BigQuery connects models to large analytical datasets that need batch processing and frequent SQL access.
Conversational data agents let users ask questions about database data in natural language instead of writing SQL. A data agent (opens in a new tab) combines a persona, a set of data sources, and business knowledge. Data flows from the user's question to the agent, the agent translates the question into SQL, the query executes against the database, and the agent returns a summary or visualization. Authored context, including schema descriptions and system instructions, is the main lever for accuracy.
Developers can also manage Spanner (opens in a new tab) from development environments without switching tools. The Data Agent Kit extension works in Visual Studio Code, the plugin works in coding agents, and the Model Context Protocol (MCP) server exposes Spanner resources to editors. The MCP configuration names the project, instance, and database through environment variables such as SPANNER_PROJECT, SPANNER_INSTANCE, and SPANNER_DATABASE.
Tools in Gemini Enterprise for Customer Experience connect virtual agents to customer systems to retrieve, update, format, or analyze information. Creating a tool follows an ordered procedure where only same-project tools can be selected:
Workflow automation with n8n on Cloud Run shows the same pattern at larger scale. Developers compose agentic workloads in a visual, node-based editor, and connected apps such as Google Workspace become tools the workflow calls during the process.
Agent Registry is the index that lets agents be discovered through their metadata and capabilities. Registration depends on where the agent runs and whether it implements the Agent2Agent (A2A) protocol, an open standard for declaring capabilities and identity. Automatic registration keeps the index in sync when agents are deployed, updated, or deleted, while external runtimes require manual registration through the Agent Registry API.
On Google Kubernetes Engine (GKE), discovery depends on labels and annotations. The registry.gke.io/functional-type: "AGENT" label identifies the deployment as an AI agent workload, and the A2A Agent Card annotation tells the controller where to find the agent's declared skills. Built-in Google agents are ingested automatically, and agents on Agent Runtime synchronize automatically on deploy, update, and delete.
Relevant Google Cloud services for agent tooling provide the compute to run agents, the storage to remember context, and the pre-built APIs to handle speech, language, and documents. The agent logic lives in a hosted environment, calls outward to these tools through APIs, and uses returned results to decide the next action.
Cloud Run (opens in a new tab) is a serverless execution environment that hosts agent logic without requiring the team to manage servers. An example is an n8n agentic workflow hosted on Cloud Run, where traffic flows from the workflow outward to each connected app or service and results flow back into the workflow. The platform provides compute when a request or event arrives and scales with demand.
Visual workflow building makes this execution model accessible without writing integration code. The n8n platform uses a visual, node-based editor where each node represents a step, an app connection, or a service call. An agent built this way can connect Google Workspace services as tools, so reasoning in the agent leads directly to actions in mail, documents, or calendar.
Agent persistence separates unstructured file storage from structured memory and state. Cloud Storage is object storage for unstructured data such as files, images, and large datasets. Database services hold structured records and vector embeddings that agents query to remember context and maintain state.
The two layers often work together in one flow. The agent pulls a raw document from Cloud Storage, sends it through a processing step, and writes the resulting embeddings or records into a database such as AlloyDB for PostgreSQL. Choosing between them depends on whether the agent is storing a blob of content or looking up a record it will filter, join, or similarity-search.
| Service | Data shape | Agent use |
|---|---|---|
| Cloud Storage | Unstructured objects such as files and datasets | Storing training data, static assets, and file-based context |
| AlloyDB for PostgreSQL | Structured records and vector embeddings | Vector store for RAG retrieval |
| Cloud SQL | Structured records and vector embeddings | RAG infrastructure on Google Kubernetes Engine |
| Spanner | Structured records across regions | Global multi-tier application state |
| BigQuery (opens in a new tab) and Bigtable | Analytical and high-throughput records | Large-scale query and lookup workloads |
Specialized AI APIs let an agent work with speech, other languages, and structured documents without model training. A developer calls the API that already understands a specific format, and the agent uses the returned result. Speech-to-Text transcribes spoken voice to text, and Text-to-Speech synthesizes written text into a human-sounding voice.
The voice APIs form a loop where user audio flows into Speech-to-Text, the resulting text flows to the agent for processing, and the agent's text reply flows into Text-to-Speech, which returns audio to the user. Text-to-Speech Custom Voice extends this by training a custom voice from studio-quality recordings for a unique brand voice. Translation translates text between languages for documents, websites, and multilingual chatbots, while Natural Language analyzes text so the agent can act on meaning.
Document AI (opens in a new tab) extracts and analyzes data from business documents such as reports, invoices, receipts, and contracts. Its output can highlight errors or compliance issues, identify risks, or uncover anomalies that indicate fraud. Document AI belongs to a family of pre-trained models that also includes Vision AI, Natural Language API, and Video Intelligence API.
Exam tip: A pre-trained API is the first option to check for classification or detection; custom training is only considered when the pre-trained model cannot meet requirements and enough training data exists.
Choosing between a pre-trained API and custom training follows an ordered path that starts with what already exists. First, the team checks whether a pre-trained model such as Document AI, Vision AI, Natural Language API, or Video Intelligence API meets the requirements. If a pre-trained model meets the requirements, the team uses it with no training work.
Only if no pre-trained model fits does the team consider custom training. Then the team checks whether enough training data is available. If sufficient data exists, the choice depends on priority: choose custom training of a traditional model when high control is required over algorithms, loss functions, explainability, or hyperparameters, and choose generative AI when faster go-to-market is the priority. Conversational platforms such as Dialogflow ES, Dialogflow CX, and Contact Center AI then call the chosen API as a tool for hearing, speaking, translating, or reading documents.
Google AI Studio and Vertex AI Studio are two different entry points to Gemini models, divided by lifecycle stage and governance needs. Google AI Studio supports fast prototyping outside a Cloud project, while Vertex AI Studio supports the full lifecycle from customization to deployment and monitoring inside Vertex AI on Google Cloud.
Google AI Studio is a browser-based tool for trying Gemini models quickly with minimal setup. A builder signs in, writes a prompt, adjusts basic settings such as temperature and output length, and receives a response within seconds. The prompt can then be turned into code that calls the Gemini API with an API key, so data flows directly from the application to the model and back.
Google AI Studio is designed for speed and simplicity rather than control. It does not manage Google Cloud projects, Identity and Access Management permissions, virtual private networks, or long-term operations. Choose Google AI Studio when the goal is to test an idea, compare prompts, explore multimodal inputs, and build a first working demo before committing to enterprise deployment.
Vertex AI Studio is the generative AI console inside Vertex AI for customizing and operating models at scale. A team works inside a Google Cloud project, so the model, data, tuning jobs, and deployed endpoints all inherit the project's ownership, billing, access controls, and network boundaries. This structure lets multiple roles collaborate securely from experimentation through production.
Vertex AI Studio supports the advanced stages of the lifecycle in order, where each stage depends on the previous one:
A common path is to start in Google AI Studio and then move to Vertex AI Studio. A prompt that works as a demo is rebuilt inside a Cloud project, where the team tunes it, evaluates it against business data, and deploys it behind a secure endpoint.
Exam tip: Use Google AI Studio for fast, API-key prototyping outside a Cloud project, and use Vertex AI Studio when you need tuning, evaluation, governed deployment, and monitoring inside Google Cloud.
Enterprise integration is the reason to choose Vertex AI Studio for production work. Vertex AI Studio runs inside a Google Cloud project with billing, enabled APIs, and Identity and Access Management (IAM) (opens in a new tab) controls, so prototypes can move to compliant deployment without leaving the enterprise boundary. A principal can be a user account or a service account for an application, and each principal receives only the roles it needs under least privilege.
Vertex AI Studio supports controls such as customer-managed encryption, network perimeters, data residency, and audit transparency. Those controls apply to the project boundary, where Vertex AI Studio sits on one side and the public internet sits on the other. Some grounding features bypass those controls when they call external services, including Grounding with Google Search (opens in a new tab), where CMEK, DRZ, VPC-SC, and AXT do not apply to the search requests.
Data integration connects Vertex AI Studio directly to operational data. Before building a data-connected agent, the project must have billing enabled and must enable the Spanner, Data Analytics API with Gemini, and Gemini for Google Cloud APIs. A conversational data agent then translates a natural-language question into SQL, executes it against Spanner, and returns results as a summary or visualization. The agent moves through ordered lifecycle states, starting as draft for personal testing and moving to published to become available to other users, through an API, or through agent interactions (opens in a new tab) in Gemini Enterprise.
Continue with this study guide: Determining when to use Vertex AI Studio and Google AI Studio
Extensions connect generative AI agents to external systems and live APIs, function calling lets an agent run a defined operation or custom code such as a calculation or workflow step, and data stores ground answers in private organizational data. Data flows from the external API or database into the model prompt, then from the model to the user as a grounded answer or action.
Retrieval-Augmented Generation (RAG) grounds responses by retrieving relevant documents, records, or API results first and adding them to the prompt so the model generates answers based on verifiable information. The user query is converted into an embedding and compared against indexed data, then the system retrieves the most relevant results and the model receives the original query plus the retrieved facts to generate the final answer.
Use Google AI Studio for fast browser-based prototyping with an API key outside a Google Cloud project to test an idea, compare prompts, explore multimodal inputs, and build a first working demo. Use Vertex AI Studio inside Vertex AI on Google Cloud when you need the full production lifecycle of prototyping, customizing with supervised fine-tuning, evaluating base and tuned models, deploying to Model Registry and Vertex AI endpoints, and monitoring quality, usage, and drift.
Choose Vertex AI Search when a ready managed search engine over many data sources is needed without managing embeddings, Vertex AI RAG Engine when a customizable retrieval pipeline with a managed RAG orchestrator is needed, and Vertex AI Vector Search when low-latency large-scale similarity search with an optimized vector database is needed. Queries and documents must use compatible embedding representations to match correctly because the vector database depends on the embedding model that created the vectors.
Last updated