Google Professional Agentic Architect (PAA) Practice Test
Design, Evaluate, and Govern AI Agents on Google Cloud.
500 scenario-based practice questions covering all 5 PAA domains: low-code agent building, coding agents, custom agents with ADK, evaluation and deployment, and security and governance. Mapped to the official exam guide.
No credit card required · Upgrade to 500 questions for $11.99
500 Exam-Style Questions
7-Day Money-Back Guarantee
Is this for you?
You have built an agent with ADK and want practice choosing between sequential, parallel and graph workflows, and between MCP and A2A, for a specific handoff
You have to pick a model under cost and security limits, such as an SLM versus an LLM or a self-hosted versus a SaaS model, and want scenarios that force the trade-off
You run agents in production and want practice choosing between Agent Runtime, Cloud Run and GKE, and diagnosing drift, tool invocation latency and reasoning loops
You work across Gemini Enterprise and Antigravity and want to test yourself on state-based flows in CX Agent Studio as well as coding agents configured with MCP servers, skills and hooks
Not for you if:
If you are new to generative AI on Google Cloud, start with the Generative AI Leader (GAIL) exam; Google recommends 1+ years building agentic solutions on Google Cloud before PAA
If you want braindumps or walkthroughs of the PAA hands-on labs: CertSharp offers neither, and its questions prepare you for the multiple-choice exam only
The Google Professional Agentic Architect (PAA) is a professional-level Google Cloud certification. The exam has 50–60 questions, runs 2 hours, costs $200, and requires a passing score of Not published by Google. CertSharp provides 500 exam-style PAA practice questions — 30 free with no credit card — each with a detailed explanation.
Using coding agents for application development (~17%)
3
Developing custom agents (~33%)
4
Evaluating and deploying agentic workflows (~22%)
5
Securing and governing agentic workflows (~15%)
From the exam guide
What the PAA exam actually tests
PAA is a design exam: questions describe an agent, its data and its constraints, then ask which model, tool, runtime or control fits best. The guide names products across Gemini Enterprise, Antigravity and Agent Platform, several of them recently renamed, so knowing what each product does matters as much as knowing agent patterns.
Low-code, coding agents or ADK
The first two sections cover state-based flows with pages, transition routes and event handlers in Agent Designer and CX Agent Studio, then coding agents in Antigravity and Claude Code on Google Cloud. You need to know when a coding agent belongs in a sandbox such as Cloud Workstations or GKE, and how skills, hooks, rules and subagents customize it.
Retrieval, memory and tool access
Custom agent questions ask you to configure embedding models, similarity scoring and reranking with Vector Search or Agent Retrieval, and to set up managed sessions and Agent Platform Memory Bank. You also decide how an agent reaches managed databases, APIs and third-party SaaS tools through Google Cloud MCP Servers and Agent Registry.
Evaluation with test sets
The guide expects you to build test sets from golden data, prompts and edge cases, and to choose between ADK evalsets, the Agent Platform GenAI evaluation service and custom autoraters. Continuous evaluation pipelines judge the agent's tool execution against success criteria, not only its final answers.
Identity, gateways and guardrails
Security questions cover OAuth 2.0 for agent-to-tool API calls, principal access boundary policies through Agent Identity, and Agent Gateway for tracking agents and propagating identity to data. You decide where Model Armor, Agent Registry policies and human-in-the-loop approval belong in a workflow.
Why CertSharp
Built for Exam Day, Not Just Study Time
Study guides teach concepts. We prepare you for the actual PAA exam.
1
Custom-agent scenarios (the largest domain at about 33%) cover ADK, sessions and memory, RAG and vector retrieval, and multiagent orchestration with MCP and A2A
2
Security and governance questions go deep on Agent Identity, Agent Gateway, Agent Registry, Model Armor, and human-in-the-loop guardrails
3
Evaluation and deployment questions test golden datasets, continuous evaluation, and choosing between Agent Runtime, Cloud Run, and GKE
4
Written against the current exam guide, with current product names such as Agent Runtime (formerly Agent Engine) and Agent Search (formerly Vertex AI Search)
Free Sample Questions
PAA Practice Questions (Free Sample)
Real PAA exam-style questions from our 500-question Google Professional Agentic Architect bank, each with the correct answer and a detailed explanation.
Question 1 — PAA
An agent needs to read customer rows from BigQuery but must not see raw values in the email and phone columns. Which control is most appropriate?
A
A system instruction telling the agent not to display the email and phone columns in responses
B
A Model Armor response filter that deletes the two columns after the query result is returned to the model
C
BigQuery column-level security with policy tags or dynamic data masking for the agent identity
D
A separate project copy of the table that includes raw values, accessed through the agent's broad reader role
Correct answer: C — Explanation
Column-level security and masking enforce restrictions at the data layer, so the agent never receives raw values. Instructions are advisory, response filtering occurs after the model has seen the data, and raw copies spread sensitive data.
Question 2Securing and governing agentic workflowsMedium
A medical agent is wrongly blocked on legitimate clinical questions by a Model Armor responsible AI filter. What is the best tuning action?
A
Lower the confidence threshold for the affected filter so more content is flagged and reviewed by hand
B
Raise the confidence threshold for the affected filter so detections must be surer, then retest
C
Disable Model Armor on the gateway for the agent and rely on the model's built-in refusals instead
D
Move the agent to Cloud Run, since runtimes apply different filter thresholds to the same template
Show answer & explanation
Correct answer: B — Raise the confidence threshold for the affected filter so detections must be surer, then retest
Confidence thresholds control how certain a detection must be to trigger, so raising the threshold reduces false positives for a specific filter. Lowering it increases blocking, disabling screening removes protection, and thresholds do not vary by runtime.
Question 3Evaluating and deploying agentic workflowsEasy
Which Cloud Monitoring capability lets a team define a target such as 99 percent of agent requests answered under three seconds and alert when the error budget burns too fast?
A
Quota dashboards, which show consumption of service limits but do not measure user-facing latency targets
B
Service level objectives with burn-rate alerting, which measure compliance against the target over a window
C
Log sinks, which route entries to other destinations but do not calculate objectives or error budgets
D
Uptime checks alone, which probe an endpoint for availability but do not track latency percentiles and error budgets
Show answer & explanation
Correct answer: B — Service level objectives with burn-rate alerting, which measure compliance against the target over a window
SLOs express reliability targets, track the remaining error budget, and support burn-rate alerts. Uptime checks, log sinks, and quota dashboards do not compute objective compliance for latency and success rates.
Question 4Developing custom agentsEasy
A prototype agent uses an in-memory session service. After a container restart during a rollout, users lose their conversations. What is the correct fix?
A
Increase the container memory limit so that the in-memory sessions are retained longer
B
Lower the model temperature so that conversations become shorter and easier to rebuild
C
Switch to a persistent session service, such as a managed or database-backed implementation
D
Store the conversation in the system prompt so that the model remembers it after restarts in production
Show answer & explanation
Correct answer: C — Switch to a persistent session service, such as a managed or database-backed implementation
In-memory sessions live in process memory and vanish on restart, so persistence needs a durable session store. More memory does not survive a restart, temperature is irrelevant, and a static prompt cannot hold per-user history.
Question 5Using coding agents for application developmentEasy
A large refactor requires the main agent to explore many modules without filling its own context with search output. Which Antigravity feature helps most?
A
A bigger editor window, which increases the amount of text that the model is able to keep in its working memory
B
A single rule that tells the agent to read each file in the repository before it begins making any change
C
Subagents that investigate modules in their own separate contexts so search output stays out of the main agent's context
D
A hook that appends the contents of the entire set of modules to the prompt before each request is sent to the model
Show answer & explanation
Correct answer: C — Subagents that investigate modules in their own separate contexts so search output stays out of the main agent's context
Subagents start with a clean slate and do not inherit the parent's conversation history, so the exploration and its search output happen in their own context windows rather than the main agent's. A parent can also run several subagents concurrently. Reading every file or appending all modules to the prompt bloats the main context, and the editor window size has no effect on model memory.
Question 6Building agents using low-code toolsHard
A designer wants a Workflow Builder (formerly Agent Designer) agent to use a company knowledge source and also follow a fixed escalation rule. Which configuration is appropriate?
A
Attach the knowledge source but remove the escalation rule, because grounded agents escalate on their own whenever answers are uncertain
B
Create the escalation rule as a data source and the knowledge as an instruction, since the two settings are interchangeable fields
C
Place the escalation rule inside the knowledge source documents and leave the instruction empty, since documents override instructions
D
Attach the knowledge source as a tool or data source and state the escalation rule in the instruction with a clear trigger condition
Show answer & explanation
Correct answer: D — Attach the knowledge source as a tool or data source and state the escalation rule in the instruction with a clear trigger condition
Knowledge belongs in connected data stores or sources that the agent is allowed to use, while behavior such as an escalation rule belongs in the agent's clear instructions with an explicit trigger. Knowledge content does not replace instructions, data and instruction settings are different fields, and grounding alone does not make an agent escalate.
Question 7Securing and governing agentic workflowsHard
An email-triage agent can call a send-payment tool. An attacker embeds instructions in an incoming email asking for a transfer. Which combination is the strongest mitigation?
A
Instruct the agent to ignore instructions found inside emails, which model compliance cannot enforce reliably
B
Screen tool responses with Model Armor, restrict the payment tool by gateway policy, require human approval
C
Encrypt the mailbox with CMEK, which protects stored data but not content the agent legitimately reads
D
Set model temperature to 0, which makes outputs more deterministic and therefore immune to injected instructions
Show answer & explanation
Correct answer: B — Screen tool responses with Model Armor, restrict the payment tool by gateway policy, require human approval
Indirect prompt injection needs layered, enforceable controls: content screening, tool-level policy, and human approval for high-impact actions. Instructions are not enforcement, CMEK does not address legitimate reads, and low temperature does not prevent following injected text.
Question 8Evaluating and deploying agentic workflowsMedium
Production quality scores for an agent fall steadily over two months while the offline evalset still passes. What is the most useful remedy?
A
Switch off the autorater, because a disagreement with the offline set means the autorater is the faulty component
B
Delete the failing production samples from dashboards, because offline results are the authoritative quality record across the suite
C
Add newly observed production failures to the golden set after expert labeling, so the benchmark tracks real traffic and drift
D
Raise the pass threshold on the existing evalset, because stricter scoring on old cases detects new failure types
Show answer & explanation
Correct answer: C — Add newly observed production failures to the golden set after expert labeling, so the benchmark tracks real traffic and drift
A static golden set can no longer represent shifted user behavior, so failing production cases should be reviewed, labeled and folded in as a feedback loop. Stricter thresholds on old cases cannot detect unseen failures, and hiding or disabling the production signal removes the evidence.
Question 9Developing custom agentsMedium
When is a graph workflow a better choice than a plain SequentialAgent for orchestrating several agents?
A
When the steps routinely run in the same fixed order and each depends solely on the previous output
B
When the path includes conditional branches, joins, or cycles that depend on intermediate results
C
When the team wants to avoid defining any control flow and let the model improvise the sequence
D
When the agents are deployed in separate regions and need to share one persistent model cache
Show answer & explanation
Correct answer: B — When the path includes conditional branches, joins, or cycles that depend on intermediate results
Graph workflows suit non-linear flows: branching on results, merging paths, or revisiting nodes. A fixed linear order is exactly what SequentialAgent handles simply. Letting the model improvise is the opposite of explicit graph control, and region placement is a deployment concern that does not decide the orchestration shape.
Question 10Using coding agents for application developmentMedium
A security reviewer is concerned that a third-party MCP server added to a coding agent could expose sensitive internal repositories. Which control best reduces the risk before adoption?
A
Give the server the developer's personal admin token so that it can authenticate without prompting and avoid any permission errors
B
Review the server's tool definitions and source, scope its credentials to least privilege, and restrict it to an approved allowlist of servers
C
Trust the server because it implements the MCP specification, since protocol compliance implies that its behavior is safe for internal data, for now
D
Disable logging for the server so that sensitive repository content does not appear in audit trails that other teams may review
Show answer & explanation
Correct answer: B — Review the server's tool definitions and source, scope its credentials to least privilege, and restrict it to an approved allowlist of servers
A third-party MCP server runs with whatever access it is given, so vetting its tools and code, scoping credentials and allowlisting are the right controls. Protocol compliance says nothing about trustworthiness. Admin tokens violate least privilege, and disabling logs removes the audit evidence needed to detect misuse.
Want more PAA practice questions?
Take the free PAA practice test: 30 questions per session with instant explanations. Unlock all 500 questions plus timed exam mode for $11.99 one-time.
Everything you need to know about Google Professional Agentic Architect
What is the Google Professional Agentic Architect certification?
The Google Professional Agentic Architect (PAA) validates that you can design and manage autonomous, AI-driven agentic workflows on Google Cloud. The multiple-choice exam costs $200, has 50-60 questions, and runs 2 hours. Passing it unlocks a second part of hands-on labs in Google Skills.
How hard is the Google PAA exam?
PAA is a professional-level exam that tests design judgment more than memorization: choosing between low-code tools, coding agents, and custom ADK agents, designing RAG and multiagent workflows, evaluating agents, and securing them. Google recommends 3+ years of hands-on cloud experience, including 1+ years building agentic solutions on Google Cloud.
Is the Google Professional Agentic Architect worth it?
It is a new credential that targets a fast-growing skill set: building and governing AI agents rather than just prompting models. It is most useful for developers and architects who already build agentic solutions and want to validate that experience with a Google Cloud credential.
What experience do I need for PAA?
Google lists no prerequisites. Its recommended experience is 3+ years of hands-on experience building, testing, deploying, and managing cloud solutions, including 1+ years building agentic solutions using Google Cloud.
What topics are on the Google PAA exam?
The five exam guide sections are building agents with low-code tools (~13%), using coding agents for application development (~17%), developing custom agents (~33%), evaluating and deploying agentic workflows (~22%), and securing and governing agentic workflows (~15%). Tools in scope include ADK, Agent Gateway, Agent Identity, Agent Registry, Agent Runtime, Model Armor, MCP, and A2A.
How many questions are on the Google PAA exam?
The multiple-choice exam has 50-60 questions and a 2-hour time limit. Google does not publish a passing score. After you pass, you receive instructions to take the separate PAA hands-on labs in Google Skills.
How long should I prepare for Google PAA?
It depends on your agent-building experience. Plan study time across all five domains, give extra time to custom agents (about 33% of the exam) and evaluation and deployment (about 22%), and build a small agent with ADK so the design tradeoffs feel concrete.
How does Google PAA compare with Professional Cloud Architect?
Professional Cloud Architect covers designing broad Google Cloud solutions across infrastructure, networking, and security. Professional Agentic Architect is specialized: it focuses on agent design patterns, ADK, RAG, MCP and A2A, agent evaluation, and agent security and governance. They validate different skills, and neither is a prerequisite for the other.
How much does the Google Professional Agentic Architect exam cost?
The registration fee is $200 USD, plus tax where applicable. The exam is 2 hours, offered in English, and can be taken with online or onsite proctoring. Google lists the certification validity period as 1 year.
Is there a free Google PAA practice test?
Yes. CertSharp offers a free Google PAA practice test for the Google Professional Agentic Architect exam: 30 free questions per session with detailed explanations, no credit card required. The full 500-question PAA bank, including timed exam mode, is $11.99 one-time with lifetime access.
Is the Google Professional Agentic Architect exam generally available?
Google ran a beta of the exam from September 8 to September 30, 2026, and lists general availability from mid-November 2026. Check the official certification page for current registration. Passing the multiple choice exam is followed by required hands-on labs in Google Skills, which Google says must be completed within two months to earn the certification.
Also on CertSharp
Related to the PAA Exam
Where PAA candidates often go next, each with its own 500-question bank.