GCP Agentic Architect Concept Map (Mind Map): Section 1

Every topic from Section 1 of the exam on one page, with the pairs the exam swaps to trick you. It's about 13% of your score.

1.1 Configuring agentic workflows and behavior using low-code tools

Workflow Builder, CX Agent Studio, flows, pages and prompts

Picking the low-code tool

Gemini Enterprise: Google’s agent workspace for employees: one chat app that searches company data, runs Google-made agents and runs agents built by the company. Admins connect the data and choose who sees which agent.
Agent Platform (formerly Vertex AI): Gemini Enterprise Agent Platform is the developer platform behind the agents: models, Agent Development Kit, Agent Runtime, evaluation and governance. It is not a no-code tool for business users.
Workflow Builder (formerly Agent Designer): The no-code and low-code builder inside Gemini Enterprise. An employee describes the agent in plain words or uses the visual editor, then tests, shares and schedules it.
CX Agent Studio (Customer Experience Agent Studio): Low-code builder for customer-facing conversational agents, on chat and voice, with telephony and contact center connections. Google presents it as the evolution of Dialogflow CX, built on the Agent Development Kit.
Agent Studio (formerly Vertex AI Studio): The Agent Platform console workspace to try models and write and test prompts. Useful to tune a prompt before you paste it into an agent.
Agent Gallery: The catalog inside Gemini Enterprise where employees find and open agents: Google-made agents and agents shared by their organization.
Agent Development Kit (ADK): Google’s open-source code framework for agents. Listed here because it is where you go when low-code tools reach their limits.

Workflow Builder building blocks

Chat agent: A conversational agent in Workflow Builder. It answers in a chat and decides each reply itself, so results are mostly non-deterministic. Can be single-step or multi-step.
Workflow: A sequence of steps in Workflow Builder that mixes AI steps and human steps, with deterministic or AI-decided logic. Can run on demand or on a schedule.
Multi-step agent (main agent and subagents): A main agent that hands parts of a task to subagents, each with its own instructions and tools. The visual editor shows how they connect.
Scheduled runs: Workflow Builder can run a workflow on a recurring schedule, for example a weekly report, without anyone starting it.
Preview and test: A live test area in the builder where you try the agent while you build it, before you share it.
Admin-provisioned data and tools: Employees can only add the data connectors, tools and permissions that the Gemini Enterprise admin has set up, such as Drive, Gmail or Calendar.
Importing ADK and A2A agents: Gemini Enterprise can import agents built in code (ADK) or reachable over A2A, so employees use them next to no-code agents and admins manage them in one place.

State-based flows

Flow-based agent: In CX Agent Studio, an agent that reuses Dialogflow CX flows. The conversation is handed to the flow until it ends, then returns to the calling agent. Best for strict, step-by-step logic.
Flow: A group of related pages that covers one topic of the conversation, such as payments or booking.
Page: One state of the conversation, like a node in a state machine. The agent stays on a page until a route or an event moves it.
Transition route: A rule that fires when the user’s request matches an intent, a condition is true, or both, and can move the conversation to another page.
Event handler: A handler that runs when an event happens: the user says nothing, nothing matches, a webhook fails, or a custom event you raise. Can sit on a flow, a page or a parameter.
Intent: What the user wants, recognized from example phrases in classic flow design. For CX Agent Studio, Google advises letting the model classify open requests and keeping flows for strict logic.
Parameters and forms: Values the page must collect, like a date or an account number. The page keeps asking until the form is complete.
Session parameters (input and output mapping): Values that travel with the conversation. In CX Agent Studio, variables are sent into a flow as session parameters, and flow parameters come back as variables.
Webhook: A call from the flow to your own backend service, for example to look up an order or save a booking.

Instruction-based agents in CX Agent Studio

Instruction-based agent: A CX Agent Studio agent driven by natural-language instructions and tools. The model decides what to say and which tool to call.
Agent application: The whole tree of agents in CX Agent Studio: one root agent plus its sub-agents, with shared settings.
Root agent: The entry point that talks to the customer first, works out the goal and passes specific work to sub-agents.
Sub-agent (transfer): A specialized agent for one task or domain. When the conversation is transferred to it, it takes over talking to the customer.
Agent as a tool: The active agent calls another agent like a tool and keeps the conversation. It can run in the background (asynchronous, the default) or wait for the answer.
Handoff rules: Rules that force or block a transfer between parent and child agents based on session variables or code. Set up in the console, no model decision involved.
Callbacks: Python code that runs before or after the agent, the model call or a tool call, to check input, change output, skip a step or redact data.
Guardrails: Built-in protections: prompt guard against attacks, blocklists of words, safety levels (relaxed, balanced, strict), custom rules and supervisor agents. Each sets an outcome: fixed reply, handoff or generated reply.
Variables: Named values the agents can read and write. In instructions, {{name}} pastes a static value into the prompt; {name} refers to a dynamic variable that tools and callbacks update.

Steering behavior with prompts

System instructions: Standing guidance the model reads on every turn: the agent’s role, persona, goal and constraints.
Structured instructions (role, persona, constraints, taskflow): CX Agent Studio recommends an XML layout with role, persona, constraints, a taskflow of steps and examples. The “Restructure instructions” button converts plain text into it.
Global instructions: Instructions set once for the whole agent application and inherited by every agent, sent with every turn.
Prompt template: A reusable prompt with blanks filled at run time from variables, so the same wording serves many cases.
Zero-shot prompt: An instruction with no examples. Try it first: clear instructions often suffice.
Few-shot examples: A few sample exchanges placed in the prompt to show the exact format or behavior wanted. CX Agent Studio advises using them sparingly.
Chain-of-thought: Asking the model to work through a problem in steps before answering, which helps on multi-step reasoning such as eligibility checks.

1.2 Connecting enterprise data to Gemini Enterprise

Connectors, data stores and multimodal content

Connecting the sources

Agent Search (formerly Vertex AI Search): Google’s managed search and retrieval service. It indexes your documents, websites and tables into data stores so agents can find answers and ground replies.
Data store: The container that holds one source’s indexed content. Gemini Enterprise creates a separate data store for each type of entity a connector brings in.
Connector: A ready-made link from Gemini Enterprise to a Google or third-party source such as Drive, Gmail, BigQuery, Jira or ServiceNow.
Federated search (data federation): Gemini Enterprise asks the source at query time. Nothing is copied, so no storage, but quality can be lower because the data is not indexed, and the query is sent to the source.
Data ingestion (indexed): The connector copies the data into the Gemini Enterprise index. Search quality is better, but it takes storage and sync time.
Full sync: Replaces the whole content of the data store with a fresh copy from the source.
Incremental sync: Brings in only added or updated items. Must run more often than the full sync (the default is every 3 hours).
Access control on sources: Gemini Enterprise keeps the source’s permissions: a user only gets results from documents they may open. Set up access control and identity before the first sync.
Identity provider: Who signs users in: Google identity, or a third-party provider through Workforce Identity Federation. Needed to match users with source permissions.
Grounding: Making the agent answer from retrieved company data, with citations, instead of from the model’s general knowledge.
Data store tool: In CX Agent Studio, a tool that lets the agent query an Agent Search data store, such as a website or files in Cloud Storage.

Unstructured and multimodal content

Unstructured data: Documents such as PDF, Word, slides, HTML, often uploaded from Cloud Storage. Needs parsing and chunking before search.
Structured data: Records with fields, for example a BigQuery table or JSON rows, searched and filtered by those fields.
Cloud Storage: Google Cloud’s object storage, the usual landing place for files, images, audio and video before they are indexed.
BigQuery: Google Cloud’s data warehouse. A source for structured data stores; Gemini Enterprise also offers a federated query mode for it (Preview in October 2026).
Digital parser: The default parser: extracts the machine-readable text of a document.
OCR parser: Reads text that is part of an image, for scanned PDFs. A lower-cost choice when only the text is needed.
Layout parser: Detects the structure of PDF, HTML, Word, PowerPoint and Excel files: headings, paragraphs, tables, lists, images. Can also run OCR. Recommended for rich documents.
Image and table annotation: With the layout parser, a written description of each image or table is stored in the chunk with the item itself, so questions about charts and tables can be answered.
Document chunking: Splitting documents into passages before indexing. Turn it on when the data store feeds RAG; layout-aware chunking keeps sections together, and ancestor headings can be added to each chunk.
Gemini multimodal input: Gemini models read text, images, audio and video directly. A workflow can pass a recording or a photo straight to the model, or turn it into text first.
Media-to-text extraction: A common pattern: Gemini writes transcripts, captions or summaries of audio, video and images, stored as searchable text with a link to the original file.
Multimodal embeddings: Number vectors that place text, images and video in the same space, so a text query can find a matching picture or clip.

Part of the Google Cloud Professional Agentic Architect Study Kit: Concept Maps, Flashcards and practice questions. Also free: the Section 1 flashcards. New to the exam? Read the GCP Agentic Architect study guide.