Prepare and test your skills
Prepare and test your skills
Evaluating external environment interaction means judging how well an agent reaches beyond its frozen training data to use live systems and current information. A language model alone cannot know today's news, the current rows in a company database, or the live state of a customer account. 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. Choosing the right tool depends on whether the agent needs live web facts, database answers, customer-system access, code-environment data, or a multi-system workflow.
Google Workspace Google Search (GWGS) (opens in a new tab) is a companion agent tool that grounds 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, plus recommended search suggestions. By pulling live web information into the prompt, GWGS lets a companion agent answer accurately about current news, global knowledge, and other dynamic subjects. This keeps conversations rooted in factual, up-to-date content instead of only static model knowledge.
GWGS carries a specific security tradeoff. It does not support Customer-Managed Encryption Keys (CMEK), Data Residency (DRZ), VPC Service Controls (VPC-SC), or Access Transparency (AXT). When the tool is enabled, those controls do not apply to the search requests it makes. Use of the tool is governed by Section 20(k) of the Google Cloud Platform Service Specific Terms.
Exam tip: Enabling GWGS means CMEK, Data Residency, VPC Service Controls, and Access Transparency do not apply to the search requests the tool makes.
Conversational data agents (opens in a new tab) let users ask questions about database data in natural language instead of writing complex SQL. A data agent combines a persona, a set of data sources, and the business knowledge critical to its purpose. The agent translates natural language into SQL, runs the query, and can then summarize or visualize the results. Agents can be drafted for personal testing or published to make them available to other users in the project or programmatically through the API.
Authored context is the main lever a creator can use to improve accuracy. Authored context includes schema descriptions, system instructions, and structured context about expected database queries. Creating data agents requires billing enabled on the Google Cloud project and the Spanner, Data Analytics API with Gemini, and Gemini for Google Cloud APIs enabled. Access is controlled by Identity and Access Management (IAM) roles matched to the task, plus database reader permissions:
| Use case | Required role |
|---|---|
| Create data agents in a project | Gemini Data Analytics Data Agent Creator (roles/geminidataanalytics.dataAgentCreator); also grants Data Agent Owner on agents you create |
| Edit, share, or delete a data agent | Gemini Data Analytics Data Agent Owner (roles/geminidataanalytics.dataAgentOwner) |
| Edit a data agent in a project | Gemini Data Analytics Data Agent Editor (roles/geminidataanalytics.dataAgentEditor) |
| View data agents in a project | Gemini Data Analytics Data Agent Viewer (roles/geminidataanalytics.dataAgentViewer) |
| Provision agents in Gemini Enterprise | Permissions to register and manage A2A agents in the Gemini Enterprise administration console |
Creating or editing a data agent also requires an IAM user or service account added to the cluster, the spanner.databaseReader role at the project level, and project-level roles for the required databases.
Gemini Enterprise for Customer Experience (opens in a new tab) tools connect customer service virtual agents to external systems so they can retrieve, update, format, or analyze information. AI coach can reuse these same tools to improve its suggestions. Creating a tool follows an ordered procedure where the console lists only tools created in the same project, so a tool from another project cannot be selected for AI coach.
Adding a tool to AI coach happens while creating AI coach instructions. The builder signs in to the Agent Assist console and clicks AI coach, then enters $ for the tool and selects one from the list. Because the list shows only same-project tools, cross-project selection is blocked by design.
Connecting agents to Spanner in development environments means managing a Spanner database and running queries without leaving the editor or coding agent. Google Cloud Data Agent Kit provides data discovery and exploration, so developers can ask questions about Spanner data in natural language and avoid switching between Spanner command-line tools and the development environment. Spanner can also be exposed through a Model Context Protocol (MCP) server, whose configuration names the project, instance, and database. The editor then exposes the server resources and skills to the agent.
| Connection surface | Works with | What it provides |
|---|---|---|
| Data Agent Kit extension | Visual Studio Code and VS Code-based IDEs | Manage Spanner and run queries from the IDE |
| Data Agent Kit plugin | Coding agents including Claude Code and Codex CLI | Manage Spanner and run queries from the coding agent |
| MCP server | Editors such as Antigravity and Windsurf | Exposes Spanner resources and skills to the editor |
An MCP server configuration for Spanner sets a command, arguments such as --prebuilt spanner or --prebuilt spanner-postgres with --stdio, and environment variables for SPANNER_PROJECT, SPANNER_INSTANCE, and SPANNER_DATABASE. The optional SPANNER_DIALECT variable accepts googlesql or postgresql, and the default is googlesql when no dialect is specified. In Antigravity, the recommended path is the built-in MCP Store where you locate Spanner, click Install, and authorize access to the Google Cloud project. In Windsurf, you open the Cascade assistant, click the MCP icon, click Configure, and add the configuration matching the Spanner dialect.
Workflow automation with n8n on Cloud Run is a way to build AI agents by connecting apps and services into automated processes. n8n is a workflow automation platform with a visual, node-based editor that lets developers compose agentic workloads with drag-and-drop instead of writing integration code. When n8n is hosted on Cloud Run, those workflows can connect services such as Google Workspace as tools that the agent calls during a process. Data therefore flows from the connected app into the n8n workflow, then from the workflow to the agent action.
Agent Registry is the index that lets agents be discovered, and registration happens automatically or manually depending on where the agent runs. The metadata and capabilities indexed for an agent depend on whether the agent implements the Agent2Agent (A2A) protocol, an open standard that lets agents declare 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.
| Runtime | How registration happens | Extra configuration |
|---|---|---|
| Agent Runtime on Gemini Enterprise Agent Platform | Automatic when the agent is deployed; updates and deletes synchronize automatically | None |
| Built-in Google agents | Automatic ingestion into the registry | None |
| Google Kubernetes Engine (GKE) | The cluster runtime controller performs an in-cluster introspection scan | Add the registry.gke.io/functional-type: "AGENT" label, plus the a2a-protocol.org/agent-card annotation for skill discovery |
| External or unsupported runtime | Manual registration through the Agent Registry API | None |
On GKE, the registry.gke.io/functional-type: "AGENT" label identifies the deployment as an AI agent workload, and GKE also accepts the older apphub.cloud.google.com/functional-type: "AGENT" annotation for backward compatibility. The A2A Agent Card annotation tells the ingestion controller where to find the agent's declared skills.
Distinguishing between agent tooling categories means identifying the distinct roles of extensions, function calling, and data stores. Extensions connect an agent to external systems and live APIs, function calling lets an agent run defined operations or custom code, and data stores ground answers in private organizational data. Each category answers a different question about where the agent's information or action comes from.
Extensions that connect to live systems are tools that let an agent act outside its own model by calling an external API. Google Cloud Data Agent Kit is an extension for Visual Studio Code and VS Code-based IDEs, and a plugin for coding agents such as Claude Code and Codex CLI, that lets a developer manage a Spanner database and run queries without leaving the editor. The Spanner MCP server (opens in a new tab) is another extension path where a configuration file names the server command, its arguments, and environment variables for the project, instance, and database. In Gemini Enterprise for Customer Experience, tools connect customer service virtual agents to external systems so they can retrieve, update, format, or analyze information, and AI coach can reuse those same tools.
Connecting an extension through Antigravity's MCP Store follows a fixed order:
Function calling for defined operations is the category where an agent invokes a defined operation or custom logic instead of only reading data. In Gemini Enterprise for Customer Experience, a tool is created by choosing a tool type and giving it a name and optional description that the agent model uses to decide when to call it. Tool names use lowercase words separated by underscores. The n8n workflow platform on Cloud Run shows the same idea at larger scale, because developers connect apps and services into automated processes in a visual editor, and those connected services become tools the agentic workload can call.
Creating a tool in Gemini Enterprise for Customer Experience follows these steps:
Function calling is the right category when the agent must do something, such as run a calculation, trigger a process, or apply custom code, rather than only retrieve information.
Data stores for private organizational data (opens in a new tab) ground an agent's answers in the organization's own records rather than general knowledge. A Gemini Data Analytics conversational data agent combines a persona, a set of data sources, and business knowledge, and users ask questions in natural language. The agent translates the request into SQL, executes the query, and can summarize or visualize the results. Authored context, including schema descriptions, system instructions, and structured context about expected queries, is the main lever for accuracy.
Access to data agents is controlled by IAM roles such as Gemini Data Analytics Data Agent Creator, Owner, Editor, and Viewer, plus Spanner database reader permissions. Grounding can also reach outside the organization through Grounding with Google Search (opens in a new tab), which grounds a companion agent's suggestions in real-time Google Search results for current news and dynamic subjects. Enabling that search grounding 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.
Choosing between extensions, functions, and data stores depends on what the agent needs to reach and whether it must retrieve information or take action:
A decision tree that branches on needing private data, a live external API, or a defined action to choose between data store, extension, and function calling.
| 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 tools |
| Function calling | Defined operations and custom code | Tools created in Gemini Enterprise for Customer Experience; n8n workflows on Cloud Run |
| Data store | Private organizational data | Gemini Data Analytics conversational data agents over Spanner |
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.
Analyzing grounding and reliability via data connectors means describing how plugins and data stores make agent outputs accurate and verifiable. Grounding is the practice of connecting an agent to verifiable, real-time information before it answers, so outputs are based on facts instead of only training data. The method that pairs grounding with generation is called retrieval-augmented generation (RAG) (opens in a new tab), where data flows from an external data store or API into the model prompt. This process improves accuracy and reduces hallucinations, which are plausible-sounding but factually incorrect answers.
How grounding and RAG improve reliability is explained by the ordered workflow that adds retrieved facts to the prompt before generation. First, the user query is converted into an embedding and compared against indexed data. Then, the system retrieves the most semantically relevant documents, records, or API results. Only after retrieval does the model receive the original query plus the retrieved facts, so the final answer is grounded in verifiable, domain-specific information.
Agents can ground on private enterprise data, on structured databases, or on public up-to-date information. Gemini models can evaluate whether internal knowledge is sufficient or whether grounding with Google Search is required for internet-fresh facts. Agents can also ground by querying a non-vector database such as Cloud SQL or Firestore and placing the query result directly into the prompt. Choose private data stores when accuracy depends on company documents, inventory, or customer data, and choose search or API grounding when accuracy depends on current real-world facts.
Embeddings (opens in a new tab) 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, which finds the most semantically relevant information even when wording differs. This comparison only works when queries and documents use compatible representations from the same embedding approach.
Many search engines store embeddings in a vector database (opens in a new tab), which is an optimal format 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 multiple vector database solutions for different use cases, so teams select a store based on latency, scale, and whether data already lives in a SQL or analytical system.
Data stores for grounded agents are the managed search and database options that supply trusted facts for RAG. Vertex AI Search is a fully managed, optimized search solution that supports most data sources for RAG without giving direct access to the underlying embeddings. Vertex AI RAG Engine is a managed orchestrator with a LlamaIndex-like interface for building a search engine for RAG. Vertex AI Vector Search is a specialized, optimized vector database for low-latency vector search and large-scale serving.
| Option | What it provides | Choose when |
|---|---|---|
| Vertex AI Search | Managed search engine over many data sources | A ready search engine is needed for RAG without managing embeddings |
| Vertex AI RAG Engine | Managed RAG orchestrator | A customizable retrieval pipeline with a LlamaIndex-like interface is needed |
| Vertex AI Vector Search | Optimized vector database | Low-latency, large-scale similarity search is needed |
Agents that already live in operational or analytical databases can ground without moving to a separate vector system. AlloyDB and Cloud SQL support working with embeddings and allow Google AI models to be used directly from the database, which fits programmatically accessed operational (OLTP) data that requires low latency. BigQuery supports connecting large language models (LLMs) to large analytical (OLAP) datasets that require batch processing and frequent SQL access by humans or scripts. This lets the agent query the system where the trusted data already lives, then use the returned rows as grounding for the prompt.
Exam tip: Choose Vertex AI Search for a fully managed search engine, Vertex AI RAG Engine for a managed RAG orchestrator, and Vertex AI Vector Search for low-latency large-scale vector serving.
Plugins, extensions, and APIs for external grounding are managed connectors that let an agent retrieve fresh facts from live online services. Vertex AI Extensions are managed connectors that let an agent create, deploy, and manage links between large language models (LLMs) and the APIs of external systems. Instead of relying only on indexed company data, the agent calls an online service API to retrieve grounding data and augment the model prompt with fresh results. Data therefore flows from the external API to the extension, then from the extension into the prompt, and finally from the model to the user as a grounded answer.
This API-based grounding complements data-store grounding and supports real-time tasks such as checking product details, locations, calendars, or other live systems. Function calling provides the structured mechanism behind many of these calls, where the model outputs which function to call and with which parameter values, receives the function result, and incorporates that result into its reasoning. Together, data stores supply trusted domain knowledge while plugins and extensions supply live external context, so agent outputs stay accurate, verifiable, and actionable.
Extensions connect agents to external systems and live APIs, function calling runs defined operations or custom code, and data stores ground answers in private organizational data. For example, the Data Agent Kit extension for VS Code and the Spanner MCP server are extensions, tools created in Gemini Enterprise for Customer Experience and n8n workflows on Cloud Run perform function calling, and Gemini Data Analytics conversational data agents over Spanner act as data stores.
Use Vertex AI Search for a fully managed search engine for retrieval-augmented generation (RAG) without managing embeddings, Vertex AI RAG Engine for a managed RAG orchestrator with a LlamaIndex-like interface, and Vertex AI Vector Search for low-latency large-scale vector serving.
Enabling Google Workspace Google Search means Customer-Managed Encryption Keys, Data Residency, VPC Service Controls, and Access Transparency do not apply to the search requests the tool makes. The tool grounds companion agent responses in real-time Google Search results built from the end-user prompt, customer-provided context, and Google Search engine data.
An enterprise is developing a generative AI agent to assist its customer support representatives. The architecture team needs to select the appropriate tooling mechanisms to enable three core capabilities:
Which agent tooling component should the enterprise configure specifically to ground the agent's responses in their private organizational documents?
Function calling
Code Interpreter
Data stores
Extensions
Last updated