Study guide for the Professional Agentic Architect exam
Google Cloud · September 2026 exam blueprint · 95 curriculum items
Start guide
0.1 Budget one agent turn
About this guide
Each bite isolates one concept or one architecture decision. Read them in order to prepare for the exam, or come back to a single bite later as a reference — the shape is the same either way, so you always know where in a lesson an answer will be.
Lessons are made to fit into the gaps between real engineering work, including the time while your coding agent is running. There is no reading-time promise attached to that, only a deliberate scope: one decision per bite.
A Lab completes when you open its official exercise — and, if it ends with a knowledge check, once you have answered that too. Anything else that carries a knowledge check — a bite, a checkpoint, a mock exam — completes when you answer every question of it; a bite without one completes when you say so. Completion only accumulates — nothing here takes it back.
Prerequisites
Outside the exam blueprint0 of 7 items completed
- 0.1 Budget one agent turnReason about context, latency, and cost within the budget of one agent turn.
- 0.2 Trace the reason–act–observe loopTrace a reason–act–observe loop from a request to a bounded outcome, including its stopping and failure conditions.
- 0.3 Separate workflow control from model decisionsSeparate deterministic workflow control from model-directed decisions.
- 0.4 Design tools as typed capability contractsDefine agent tools with schemas and reason about their arguments, results, failures, and side effects.
- 0.5 Place information in the right state layerDistinguish turn state, session state, long-term memory, and external knowledge.
- 0.6 Connect embeddings, retrieval, grounding, and RAGExplain embeddings, vector retrieval, grounding, and the purpose of retrieval-augmented generation (RAG).
- 0.7 Separate identity from permission and delegationDistinguish authentication, authorization, delegation, workload identity, and identity propagation.
Building agents using low-code tools
13% of the exam0 of 13 items completed
Configuring agentic workflows and behavior using low-code tools
- 1.1.1 Model state with pagesModel conversational and operational state with pages in a state-based agent.
- 1.1.2 Control movement with routes and conditionsControl workflow movement with transition routes and conditions.
- 1.1.3 Handle events without creating loopsHandle no-match, no-input, tool failures, and other events without creating loops.
- 1.1.4 Choose a low-code agent builderChoose between Gemini Enterprise Agent Designer and CX Agent Studio from workflow and integration requirements.
- 1.1.5 Write effective system instructionsWrite system instructions that establish an agent's role, constraints, tool-use boundaries, and output requirements.
- 1.1.6 Correct recurring behavior with few-shot examplesUse configured few-shot examples to correct recurring behavior and tool-selection errors.
- 1.1.7 Use reasoning prompts without exposing hidden reasoningUse chain-of-thought-style prompting appropriately while recognizing its limitations and security implications.
Connecting enterprise data to Gemini Enterprise
- 1.2.1 Connect and index enterprise sourcesConnect and index structured and unstructured enterprise sources for Agent Search.
- 1.2.2 Preserve source access controls in retrievalPreserve enterprise access controls when retrieving proprietary data.
- 1.2.3 Tune grounding thresholds for safe coverageTune grounding thresholds to balance answer coverage, clarification, abstention, and hallucination risk.
- 1.2.4 Choose native multimodal processing or preprocessingUse Gemini with images, audio, and video directly, and add preprocessing only when a specific requirement demands it.
- LABBuild Agent Search Apps with AI ApplicationsBuild search apps over website, structured, and unstructured enterprise data.
- LABIntroduction to Gemini EnterpriseConnect enterprise data stores and explore grounded assistance in Gemini Enterprise.
Using coding agents for application development
17% of the exam0 of 11 items completed
Using coding agents effectively
- 2.1.1 Attach MCP servers with a minimal tool surfaceAttach an MCP server to a coding agent while exposing only the tools required for its task.
- 2.1.2 Encode engineering knowledge in reusable skillsLearn how to turn organization-specific engineering practice into focused, discoverable, tested coding-agent skills.
- 2.1.3 Choose a coding-agent sandboxChoose Antigravity, Cloud Workstations, or GKE Agent Sandbox from the execution model, isolation boundary, and operating needs.
- 2.1.4 Bound generated code by its trust levelDesign a least-privilege capability envelope for generated code across filesystem, network, secret, and IAM access.
- 2.1.5 Refactor and optimize with repository evidenceUse repository context, tests, profiles, measurements, and focused diffs to make evidence-backed application improvements.
- 2.1.6 Patch application vulnerabilities with evidenceApply an inspect–change–verify workflow that fixes an application vulnerability without hiding it or breaking intended behavior.
- LABGet Started with Vibe Coding and Antigravity CLIPractise configuring a coding agent, its tools, shell access, and an MCP server.
Customizing coding agents for enterprise workflows
- 2.2.1 Choose the right coding-agent customizationDistinguish policy, capability, interception, and delegation mechanisms when customizing a coding agent.
- 2.2.2 Run mandatory checks with hooksChoose and place hooks so required checks run at the correct boundary around an agent action.
- 2.2.3 Scope subagents with a delegation envelopeDesign bounded subagents by aligning their task, context, tools, and privileges.
- 2.2.4 Drive Agents CLI with an agent or a humanChoose who drives Agents CLI at each lifecycle phase without confusing interaction style with governance.
Developing custom agents
33% of the exam0 of 35 items completed
Designing and building agentic workflows in code
- 3.1.1 Choose between an LLM and an SLMChoose the smallest language model that meets reasoning quality, latency, throughput, and cost requirements.
- 3.1.2 Evaluate model sourcing optionsEvaluate model sourcing across hosting, licensing, control, operations, cost, and quality requirements.
- 3.1.3 Discover and evaluate models with Model GardenUse Model Garden to shortlist models and validate them against explicit architecture requirements.
- 3.1.4 Assemble an ADK agentCompose an ADK agent from a model, instructions, tools, and focused subagents.
- 3.1.5 Understand ADK executionUnderstand how runners, events, callbacks, and call limits control an ADK invocation.
- 3.1.6 Place state at the right lifetime boundaryDistinguish invocation state, persisted session context, and file-like application artifacts.
- 3.1.7 Choose who manages session persistenceChoose managed sessions or application-managed persistence by separating the session contract from its storage implementation.
- 3.1.8 Decide what to remember across sessionsDecide which information belongs in long-term Memory Bank and when a later session should retrieve it.
- 3.1.9 Configure Agents CLI capabilities and operating modeConfigure Agents CLI skills and plugin packaging, then choose whether a coding agent or a human should drive the CLI.
- 3.1.10 Distinguish Skill Registry from Agent RegistryChoose the registry that matches reusable capability distribution or deployed-agent discovery and governance.
- LABGet Started with Agent Development Kit (ADK)Build and inspect an ADK agent with tools, sessions, and multiple execution surfaces.
- LABIntroduction to Gemini 3Exercise Gemini model controls, multimodal input, structured output, and tool use.
- LABExplore and Evaluate Models using Model GardenCompare candidate models and exercise prompt design in Model Garden.
- LABExploring Memorystore for Redis ClusterOperate a managed Redis cluster and observe persistence and scaling behaviour.
Integrating enterprise domain knowledge
- 3.2.1 Design the RAG lifecycleDesign a retrieval-augmented generation (RAG) lifecycle from source ingestion and chunking through retrieval and grounded generation.
- 3.2.2 Choose corpus embeddings, chunks, and refreshChoose compatible embedding models, chunk boundaries, and refresh policies for a retrieval corpus.
- 3.2.3 Tune retrieval precision and recallTune similarity scoring, top-k, metadata filters, and thresholds to balance relevant coverage against retrieval noise.
- 3.2.4 Apply reranking after high-recall retrievalDecide when and how to rerank retrieved candidates so that the best evidence reaches the model context.
- 3.2.5 Choose a retrieval ownership modelChoose Agent Retrieval, Vector Search, RAG Engine, or application-managed retrieval for your workload.
- 3.2.6 Balance the retrieval operating pointBalance retrieval quality, freshness, latency, scalability, and cost by setting measurable constraints and testing them together.
- 3.2.7 Give each agent a short-lived workload identityApply SPIFFE-based Agent Identity with X.509 certificates and mTLS, and distinguish it from shared, long-lived service account keys.
- 3.2.8 Expose enterprise systems through governed MCP boundariesChoose and govern MCP integration boundaries for databases, Google Cloud services, APIs, SaaS systems, and remote services.
- LABGetting Started with Vector Search and EmbeddingsGenerate embeddings, deploy a vector index, and run similarity queries.
- LABBigQuery: Qwik Start - ConsoleCreate, load, and query BigQuery datasets as an enterprise knowledge source.
- LABCloud SQL for PostgreSQL: Qwik StartCreate and connect to a managed PostgreSQL instance and perform basic SQL operations.
- LABImport Data to a Firestore DatabasePrepare and import application data into Firestore through a serverless workflow.
- CHECKCheckpoint: integrating enterprise domain knowledgeCheck the retrieval, freshness, identity, and integration-boundary decisions of 3.2 across its lessons.
Orchestrating and coordinating agentic workflows
- 3.3.1 Choose between MCP and A2AChoose MCP for bounded tool use and A2A for collaboration with an autonomous agent.
- 3.3.2 Recognize the A2A interaction lifecycleRecognize how Agent Cards, task states, status updates, and artifacts describe a remote-agent interaction.
- 3.3.3 Build sequential workflows for dependent stepsBuild fixed-order agent workflows when each step depends on a valid result from the preceding step.
- 3.3.4 Build parallel fan-out/fan-in workflowsBuild fixed parallel workflows that run independent subtasks concurrently and combine their results safely.
- 3.3.5 Build graph workflows with explicit control flowBuild graph workflows that enforce conditional branches, synchronization joins, and permitted execution paths.
- 3.3.6 Choose dynamic routing or explicit orchestrationChoose who controls each workflow transition according to the required control-flow determinism.
- 3.3.7 Coordinate the five handoff responsibilitiesCoordinate identity, discovery, execution, continuity, and policy without conflating their responsibilities during agent handoffs.
- LABConnect to Remote Agents with ADK and the Agent2Agent (A2A) SDKPublish an ADK agent over A2A and call it from another agent.
Evaluating and deploying agentic workflows
22% of the exam0 of 17 items completed
Evaluating agents in development and in production
- 4.1.1 Build representative and adversarial evaluation setsBuild evaluation cases that cover real workloads, hostile inputs, expected outcomes, tool behavior, and important edge conditions.
- 4.1.2 Distinguish agent evaluation layersDistinguish response, retrieval, trajectory, and tool-invocation evaluation by the artifact and failure each one measures.
- 4.1.3 Define measurable success criteria for agent toolsDefine testable criteria for tool selection, arguments, execution, and end-to-end task completion.
- 4.1.4 Choose the right agent evaluation mechanismChoose ADK evalsets, Agent Platform GenAI Evaluation, or custom autoraters from the evidence and judgment the evaluation requires.
- 4.1.5 Combine deterministic metrics with calibrated autoratersCombine exact checks with human-calibrated model judgments without letting one score hide critical failures.
- 4.1.6 Gate releases with continuous agent evaluationDetect agent regressions continuously and promote only candidates that satisfy explicit evaluation gates.
- LABEvaluate ADK agent performance using the Agent Platform Generative AI Evaluation ServiceEvaluate an ADK agent across tool use, trajectories, responses, and custom metrics.
Deploying and scaling production workloads
- 4.2.1 Choose a production agent runtimeChoose Agent Runtime, Cloud Run, or GKE by matching workload constraints to the right level of platform control.
- 4.2.2 Size an agent runtime for production loadSize runtime concurrency and capacity while accounting for autoscaling behavior and startup latency.
- 4.2.3 Detect and isolate agent driftDetect declining agent behavior and separate changes in models, prompts, tools, retrieval, and traffic.
- 4.2.4 Decompose end-to-end agent latencyAttribute an agent turn's latency to model, retrieval, orchestration, and tool work without double-counting parallel operations.
- 4.2.5 Bound reasoning loops with conditions and budgetsDesign reasoning loops that stop on verified outcomes or before they exceed a deterministic execution budget.
- 4.2.6 Diagnose failures with traces and logsClassify failed agent runs by finding whether the first causal fault occurred in agent logic, a tool, or infrastructure.
- 4.2.7 Balance production objectivesOptimize an agent as a constrained system in which quality, performance, reliability, hallucination rate, and cost compete.
- LABDeploy an Agent with Agent Development Kit (ADK): Challenge LabBuild a multi-agent application and deploy it to Agent Runtime.
- LABBuild an AI Agent and Configure an MCP Server on Cloud RunPackage an ADK agent and deploy its application and MCP integration on Cloud Run.
- LABUsing Cloud Trace on Kubernetes EngineDeploy a traced service on GKE and use distributed traces to diagnose latency.
Securing and governing agentic workflows
15% of the exam0 of 12 items completed
Configuring agent security and governance
- 5.1.1 Delegate user access with Auth ManagerUse Auth Manager for end-user OAuth delegation and choose correctly between user and workload authority.
- 5.1.2 Cap agent reach with Principal Access BoundariesApply Principal Access Boundary policies as a resource ceiling for an Agent Identity.
- 5.1.3 Control agent traffic with Agent GatewayUse Agent Gateway as an inline boundary for routing, authentication, policy enforcement, and correlated observability.
- 5.1.4 Govern discovery with Agent RegistryUse Agent Registry as an owned inventory of capabilities while keeping runtime authorization separate.
- 5.1.5 Compose governance across agent boundariesDesign layered governance so identity, registry, gateway, tool, and content controls enforce distinct parts of one policy.
Implementing secure agent behavior and execution
- 5.2.1 Threat-model six agent surfacesDecompose an agent threat model across prompts, model outputs, tools, data, identities, and consequential side effects.
- 5.2.2 Choose guardrails by the kind of judgmentChoose deterministic guardrails for hard policy constraints and model-based checks for judgments that require understanding meaning or intent.
- 5.2.3 Gate consequential actions with human approvalInsert a human-in-the-loop approval checkpoint before an agent performs a high-impact or irreversible action.
- 5.2.4 Propagate identity without expanding privilegePropagate user and agent identity across agent, tool, and data boundaries while keeping every action within the authority granted for the task.
- 5.2.5 Configure Model Armor runtime controlsConfigure Model Armor to inspect prompts and model responses, tune confidence thresholds, and choose between observation and active blocking.
- 5.2.6 Inspect and de-identify sensitive data safelyUse Sensitive Data Protection to inspect and de-identify data, then choose a safe inspect, redact, block, or continue response.
- LABSanitize Prompts and Responses with Model ArmorConfigure Model Armor and inspect prompts and responses for runtime threats and sensitive data.