Course Introduction
Professional certification
Google Cloud Certified - Professional Agentic Architect
Design and manage autonomous, AI-driven agentic workflows on Google Cloud. Use this guide to master ADK, A2A, MCP, RAG, evaluation, and governance — then drill flashcards and decision scenarios before exam day.
Want to pass faster?
Get practice tests + the latest deals/discount codes
Outcomes
Design autonomous AI-driven agentic workflows on Google Cloud.
Practice
Decision scenarios for low-code vs code, RAG, runtimes, and protocols.
Exam focus
Flashcards covering ADK, A2A, MCP, Agent Identity, and deployment.
Next steps
Apply patterns to evaluation, deployment, and governance.
Exam Guide + Learning Objectives
The exam validates your ability to build agentic solutions with LLMs, agent design patterns, code, and data integrations on Google Cloud — balancing reliability, performance, cost, security, and scalability.
Exam Overview
Length: 3 hours
Format: Two parts — a proctored multiple-choice exam (~80 questions, English, online or onsite) plus hands-on labs
Beta status: In beta as of September 2026 — beta fee $120 USD vs $200 retail. Google does not publish a passing score.
Validity: 1 year
Prereqs: None officially; 3+ years building cloud solutions including 1+ year building agentic solutions on Google Cloud recommended
Official page: cloud.google.com/learn/certification/agentic-architect
The 5 Exam Sections
Open each domain to review key objectives.
1) Building agents using low-code tools (~13%)
State-based workflows: pages, transition routes, and event handlers in Gemini Enterprise Agent Designer and CX Agent Studio.
Prompt design: system instructions and in-console prompt templates (few-shot, chain-of-thought).
Enterprise data: secure connections via Agent Search; multimodal ingestion (video, audio, images).
2) Using coding agents for application development (~17%)
Effective use: MCP servers, custom skills, tool access (Antigravity, Claude Code on Google Cloud).
Sandboxes: GKE, Cloud Workstations, Antigravity sandboxes.
Enterprise workflows: skills, plugins, extensions, hooks, rules, subagents; Agents CLI to build, scale, govern, optimize.
3) Developing custom agents (~33%)
Model selection: LLM vs SLM, self-hosted vs SaaS, OSS vs proprietary (cost, security, architecture).
Build: ADK; sessions and memory (Memory Bank, managed sessions).
Knowledge: RAG pipelines — embeddings, similarity scoring, reranking; Vector Search, Agent Retrieval.
Orchestration: MCP and A2A; sequential, parallel, graph workflows; Agent Runtime, agent policies.
4) Evaluating and deploying agentic workflows (~22%)
Evaluation: golden data, prompts, edge cases; ADK evalset; Gen AI evaluation service; custom autoraters; continuous pipelines.
Deployment: Agent Runtime vs Cloud Run vs GKE — match to use case, requirements, cost.
Operations: troubleshoot drift, tool latency, reasoning loops; monitor with Cloud Logging and Cloud Trace.
5) Securing and governing agentic workflows (~15%)
Security: OAuth 2.0 agent-to-tool auth (Auth Manager); principal access boundary (PAB) via Agent Identity.
Visibility: Agent Gateway monitors traffic and tracks agents; identity propagation.
Governance: Agent Registry and Model Armor; guardrails with human-in-the-loop.
Deep Dive: Critical Agentic Topics
ADK essentials
Agents: an agent is an LLM plus instructions plus tools; ADK is the open-source kit for building them in code.
Tools: typed functions the agent calls — keep them narrow, idempotent, and well-described.
Sessions: managed sessions hold conversation state across turns.
Evaluation: ADK ships evalset tooling — run golden datasets before every release.
A2A vs MCP — never mix them up
A2A (Agent2Agent): protocol for agent-to-agent communication — handoffs, delegation, multi-agent workflows.
MCP (Model Context Protocol): protocol for agent-to-tool and agent-to-data connections via MCP servers.
Google Cloud MCP servers: managed servers that connect agents to databases, APIs, and third-party SaaS.
Exam trap: if the scenario is agents talking to agents, the answer is A2A — not MCP.
RAG stack
Pipeline: chunk and embed content, retrieve with similarity scoring, rerank for relevance, then generate the grounded answer.
Vector Search: large-scale approximate nearest-neighbor retrieval over embeddings.
Agent Retrieval: agent-native managed retrieval; RAG Engine is the end-to-end managed RAG pipeline option.
Evaluate retrieval separately from response quality — bad retrieval caps answer quality.
Runtime choices
Agent Runtime: fully managed agent hosting with sessions, memory, and evaluation hooks (formerly Agent Engine).
Cloud Run: serverless containers for simple stateless agent endpoints; scales to zero.
GKE: full control — custom networking, GPUs, sandboxing for coding agents; you manage the cluster.
Rule: match runtime to requirements and cost, not hype — managed first when ops overhead matters.
Identity and governance
Agent Identity: every agent gets an identity; scope it with principal access boundary (PAB) policies.
Agent Gateway: central place to monitor agent traffic, track agents, and propagate identity.
Agent Registry: catalog of approved agents and capabilities — governance against shadow agents.
Model Armor: screens prompts and responses against safety policy; pair with human-in-the-loop for risky actions.
Exam-Day Tips
Two-part exam: proctored questions plus hands-on labs.
Pace the two parts
Part 1 is ~80 multiple-choice questions in 3 hours; part 2 is hands-on labs. Budget time so labs are not rushed.
Expect current product names
The exam is in beta and products are new: Agent Runtime (formerly Agent Engine), Agent Search (formerly Vertex AI Search). Answer with the current names.
No published passing score
Google does not publish one — answer every question and do not leave blanks.
Match managed to requirements
For "best" questions, pick the simplest managed Google Cloud service that meets the requirement with least privilege.
Read A2A vs MCP twice
The exam tests the distinction: A2A is agent-to-agent, MCP is agent-to-tool. Slow down on protocol questions.
Labs reward practice
Build an ADK agent and run an evalset against golden data before exam day so the lab mechanics feel familiar.
Strategic Golden Rules
Managed first: choose Agent Runtime over self-hosted when minimizing ops overhead.
Identity everywhere: every agent gets an identity scoped by principal access boundary (PAB) policies — never broad permissions.
Evaluate before deploy: golden data plus evalset in dev; continuous evaluation pipelines in production.
Protocols matter: A2A for agent-to-agent, MCP for agent-to-tool — the exam tests the difference.
Guardrails by default: Model Armor screening plus human-in-the-loop for high-risk or irreversible actions.
Key Questions
Reveal the answer you should give on the exam.
ADK vs Agent Designer?
ADK is the code-first open-source kit for custom agents; Agent Designer is the low-code builder inside Gemini Enterprise.
A2A vs MCP?
A2A is the protocol for agent-to-agent communication; MCP is the protocol for agent-to-tool and agent-to-data connections.
Where does long-term agent memory live?
Agent Platform Memory Bank; use managed sessions for conversation state across turns.
What does reranking do in RAG?
Reorders retrieved chunks by relevance after similarity search to improve answer quality.
Agent Runtime vs Cloud Run?
Agent Runtime is managed hosting purpose-built for agents (sessions, memory, eval hooks); Cloud Run is generic serverless containers.
What is a principal access boundary (PAB)?
PAB policies configured with Agent Identity limit which resources an agent identity can access — least privilege for agents.
What is an evalset?
ADK evaluation tooling: test sets of golden data, prompts, and edge cases you run the agent against.
What does Model Armor do?
Screens prompts and responses against safety and policy rules — a guardrail layer for agentic workflows.
When is human-in-the-loop required?
Before high-risk or irreversible agent actions — a governance guardrail, not a performance optimization.
Why is the exam two parts?
It tests both design knowledge (proctored multiple-choice, ~80 questions) and hands-on skill (labs).
Vocabulary
ADK
Agent Development Kit: open-source framework for building custom agents in code.
A2A
Agent2Agent: protocol for agent-to-agent communication and handoffs.
MCP
Model Context Protocol: standard for agent-to-tool and agent-to-data connections via MCP servers.
RAG
Retrieval-augmented generation: grounding LLM answers in enterprise data via retrieval.
Agent Runtime
Managed hosting for production agents, with sessions and memory (formerly Agent Engine).
Agent Identity
Identity and permissions for agents; scoped with principal access boundary (PAB) policies.
Agent Gateway
Central point to monitor agent traffic, track agents, and propagate identity.
Model Armor
Safety screening and policy enforcement for agent prompts and responses.
evalset
ADK test sets — golden data, prompts, and edge cases used to evaluate agents.
HITL
Human-in-the-loop: required human approval before high-risk agent actions.
Flashcards
Agent concepts, protocols, RAG, evaluation, and governance
Question Text
Click to reveal answerAnswer Text
Agentic Architecture Decision Scenarios
Click a scenario to expand the options table and the exam rule of thumb.
1. Low-code builder vs custom agent (ADK)
Scenario: The business wants an assistant over internal documents. The flow is structured (greeting, authenticate, search docs, escalate) but the team also wants custom logic later.
| Option | Pick when | Watch out |
|---|---|---|
| Gemini Enterprise Agent Designer | Structured flows, pages/routes suit the task, speed matters | Custom logic beyond state machine needs ADK anyway |
| Customer Experience Agent Studio | Customer-service conversations with human handoff | Design escalation paths up front |
| Custom agent with ADK | Complex logic, custom tools, multi-agent needs | You own code, tests, deployment, evals |
Rule of thumb: Structured, repeatable process → low-code first. Custom logic or multi-agent orchestration → ADK.
2. A2A vs MCP
Scenario: An agent must delegate research to a specialist agent, then write the findings to the CRM. Which protocol covers each leg?
| Leg | Protocol | Why |
|---|---|---|
| Agent delegates to research agent | A2A | Agent-to-agent communication and handoff |
| Agent writes to CRM | MCP | Agent-to-tool/data connection via MCP server |
Rule of thumb: Agents talking to agents = A2A. Agents calling tools or data = MCP. Never mix them up on the exam.
3. Retrieval approach for grounding
Scenario: The agent must answer from 2 million support articles with per-user access filtering and minimal ops overhead.
| Option | Pick when | Watch out |
|---|---|---|
| RAG Engine | End-to-end managed RAG pipeline wanted | Less control over internals |
| Agent Retrieval | Agent-native managed retrieval | Newer surface; confirm feature fit |
| Custom RAG on Vector Search | Custom chunking/reranking control needed at scale | You own pipeline, refresh, evals |
Rule of thumb: Managed retrieval first when ops overhead is a constraint; go custom only when you can name the specific control you need.
4. Runtime: Agent Runtime vs Cloud Run vs GKE
Scenario: Ship a production agent with conversation sessions, long-term memory, and eval hooks; traffic is steady during business hours.
| Runtime | Pick when | Watch out |
|---|---|---|
| Agent Runtime | Need managed sessions, Memory Bank, evaluation hooks | Less low-level control |
| Cloud Run | Simple stateless endpoint, spiky traffic, scale-to-zero | You wire state/sessions yourself |
| GKE | Custom networking, GPUs, sandboxes, full control | You manage the cluster and its cost |
Rule of thumb: Match runtime to requirements and cost: managed sessions/memory → Agent Runtime; bursty HTTP → Cloud Run; full control → GKE.
5. Model selection: LLM vs SLM
Scenario: High-volume ticket classification: 10M requests/month, narrow task, strict cost target, data must stay in-region.
| Option | Pick when | Watch out |
|---|---|---|
| Fine-tuned SLM | Narrow high-volume task, cost/latency sensitive | Needs quality training data and evals |
| SaaS LLM (Gemini API) | Broad reasoning, zero ops, latest capability | Data leaves your VPC; per-token cost at scale |
| Self-hosted model | Strict residency/air-gap, no external calls | You own GPUs, scaling, patching, evals |
Rule of thumb: High volume + narrow task → SLM. Cutting-edge reasoning → SaaS LLM. Residency/air-gap → self-hosted.
6. Workflow pattern: sequential vs parallel vs graph
Scenario: A research agent must gather data from five independent sources, then synthesize, then route to a specialist based on topic.
| Pattern | Shape | Use when |
|---|---|---|
| Parallel | Fan-out / fan-in | Independent subtasks: the five sources |
| Sequential | A → B → C | Dependent steps: gather then synthesize |
| Graph | Conditional routing | Dynamic routing to the specialist by topic |
Rule of thumb: Independent work fans out in parallel; dependent steps run sequential; dynamic routing needs a graph.
7. Evaluation strategy
Scenario: Before launch, the team must prove the agent handles edge cases; after launch, they must catch drift. Which evaluation where?
| Stage | Tool | Why |
|---|---|---|
| Pre-launch regression | ADK evalset | Golden data, prompts, edge cases on every change |
| Standardized managed evals | Gen AI evaluation service | Less custom code to maintain |
| Production drift detection | Continuous evaluation pipelines | Alert on metric regressions like SLOs |
| Domain-specific quality | Custom autoraters | LLM judges calibrated against human labels |
Rule of thumb: Evalset in the dev loop, continuous pipelines in production, custom autoraters for domain-specific quality.
8. Securing the agent fleet
Scenario: Agents act on behalf of users against production tools. Design authentication, authorization, and guardrails.
| Control | What it does | Exam signal |
|---|---|---|
| OAuth 2.0 + Auth Manager | Agent-to-tool auth with delegated user context | Never shared API keys; short-lived scoped tokens |
| Agent Identity + PAB | Per-agent identity bounded by principal access boundary policies | Least privilege for agents |
| Agent Gateway | Monitors traffic, tracks agents, propagates identity | Central visibility and auditability |
| Model Armor + HITL | Screens prompts/responses; human approves risky actions | Guardrails for high-risk or irreversible actions |
Rule of thumb: Authenticate with OAuth 2.0, authorize with Agent Identity + PAB, observe via the gateway, gate risky actions with HITL.