The Google Cloud Certified Professional Agentic Architect credential is Google's certification for people who design, build and run AI agents on Google Cloud. Google describes the candidate as an experienced developer or architect who builds agentic solutions while weighing reliability, performance, cost, security and scalability, with deep experience using large language models, applying agent design patterns, writing code and integrating data sources. Most questions read like a design review or an incident report: a config, an error or a log line, and four ways to respond.
The timing matters. The beta ran in September 2026, registration for the generally available exam opens on November 2, and the exam itself becomes available in mid-November. At the same time, much of the platform it tests was renamed in 2026, when Vertex AI became Gemini Enterprise Agent Platform. This guide covers how the certification works, what each section tests, which names changed, who it suits and how to prepare in about six weeks.
Exam Quick Facts
Find out where you stand first
Take the free Professional Agentic Architect practice questions cold, with no signup, and read on with your gaps in mind. For full rehearsal, preporato.com has a timed Professional Agentic Architect practice exam that follows Google's exam guide, with an explanation for every option.
How the certification works
The certification has two parts, and you need both. First comes a proctored multiple-choice exam delivered by Pearson, online or at a test center. It tests conceptual knowledge, design choices and architectural standards. If you pass it, you get access to a separate set of hands-on labs in Google Skills that validate execution and coding, and you have two months to complete them. Only after passing both parts do you receive the credential and badge.
Google says the generally available exam follows its standard professional format: a 2-hour session, with official results within 7 to 10 business days. Google has not published the question count or the passing score for it. The beta was longer, with about 80 multiple-choice questions in 3 hours.
The certification is valid for one year, half the usual term for Google Cloud professional certifications. Google's reason is the pace of change in agent technology, and it describes a continuous education path for keeping the certification active. The generally available exam costs 200 US dollars plus tax. Google lists English as the exam language, and the beta was English only.
Try PAA practice questions free
They come from the PAA practice tests, and each option shows its explanation once you answer.
- 1 timed practice exams
- Every answer explained
- A score for each domain
- $19.99 one-time
How the questions are built
Expect scenarios with evidence. A question shows a rules file that never loads, a monitor that produces no scores, or a gateway configuration that fails, and asks for the cause or for the change that meets every requirement. The wrong options are rarely absurd. They are real features that would work in a slightly different situation, so the question tests whether you know exactly what each feature can see and when it runs.
Three habits separate passing scores from near misses. Find every requirement in the scenario before reading the options, because the wrong options usually meet all but one. Read configs and log lines closely, because the cause is often a single field or default. And know the boundaries between similar controls: IAM policies decide who can reach what, a Principal Access Boundary caps what an agent identity can ever reach, Model Armor screens content, and Semantic Governance checks whether a tool call matches what the user asked.
The five sections
Google's exam guide splits the exam into five sections. Developing custom agents is a third of the exam on its own, and evaluation and deployment add more than a fifth, so code-level agent work dominates. The low-code and coding-agent sections are smaller and full of specific products that are easy to mix up under time pressure.
Core Topics
- •State-based workflows in Gemini Enterprise and CX Agent Studio
- •System instructions and prompt templates such as few-shot and chain-of-thought
- •Connecting enterprise data to Gemini Enterprise and Agent Search
- •Ingesting unstructured multimodal data such as video, audio and images
Skills Tested
Example Question Topics
- A callback cannot reach a service on a private IP. What should handle the call?
- A removed employee still sees restricted pages two days later. Which sync is missing?
Names that changed
Much of the study material online predates the April 2026 renaming, and older names still appear in documentation and in some answer options. Learn both, so an older term never throws you.
Older names and the names the exam uses
| Older name | Current name | What it is |
|---|---|---|
| Vertex AI | Gemini Enterprise Agent Platform | The platform for building, scaling, governing and optimizing agents |
| Vertex AI Agent Engine | Agent Runtime | Managed runtime for deploying and scaling agents; its API resource is still called ReasoningEngine |
| Vertex AI Search | Agent Search | Search and retrieval over enterprise data |
| Vertex AI Vector Search 2.0 | Agent Retrieval | Managed retrieval built on collections of data objects |
| Agent Engine Sessions and Memory Bank | Agent Platform Sessions and Memory Bank | Conversation history and long-term memory for agents |
| Vertex AI Studio | Agent Studio | Low-code design and prompting in the console |
| Dialogflow CX | CX Agent Studio, its successor | Customer experience agent builder, built on ADK; existing Dialogflow CX flows can be imported |
Some names moved even after Google wrote the exam guide. The Gemini Enterprise page for Agent Designer, which the guide names, is now titled Workflow Builder, and covers both chat agents and workflows. Check product names against current documentation as you study.
Who the certification is for
Google recommends three or more years of hands-on experience building, testing, deploying and managing cloud solutions, including at least one year building agentic solutions on Google Cloud. There is no prerequisite certification. In practice it suits engineers and architects who already ship on Google Cloud and now own agents end to end: the model, the framework, the data behind it, the evaluation suite and the controls around it.
It is a harder start if you have never written an agent in code. A third of the exam is ADK, retrieval and orchestration, and many questions show code or configuration. If your work with Google Cloud AI is mostly strategy and adoption, the Generative AI Leader certification is the closer fit.
Professional Agentic Architect and related Google Cloud certifications
| Certification | Built for | Choose it when |
|---|---|---|
| Professional Agentic Architect | Engineers and architects who build, evaluate, deploy and secure AI agents | You design agent systems and write the code and configuration behind them |
| Generative AI Leader | Business and technical leaders who guide generative AI adoption | You shape AI strategy and decisions without building agents yourself |
| Professional Machine Learning Engineer | Engineers who build, train and operate ML models | Your work centers on models, training pipelines and MLOps |
| Professional Cloud Architect | Architects who design Google Cloud solutions in general | You design cloud infrastructure across workloads, with agents as one part |
Claude Code is on the exam
One detail matters to anyone who already works with Anthropic's tools. The section on coding agents names Claude Code on Google Cloud next to Antigravity. Expect questions on configuring coding agents with MCP servers, skills and tools, and on running them in secure sandboxes. Claude Code can use Claude models served from your own Google Cloud project, and organizations can put the Claude apps gateway on Cloud Run in front of it for sign-in, spend limits and usage attribution.
Three Professional Agentic Architect questions
Section: Using coding agents for application development
A platform team uses Antigravity on a monorepo where Protobuf files sit in 30 service directories next to Go and Python code. Its Protobuf conventions must load every time the agent works on a .proto file, stay out of the context while the agent edits Go or Python files, and not depend on the model judging them relevant. How should the team define the conventions?
A glob rule activates when the agent works with files that match its patterns, so the conventions load for .proto work only. A model_decision rule and a skill both leave activation to the model's judgment, and an AGENTS.md file is always on for its whole directory, so it would load during Go and Python work in the same folders.
Section: Evaluating and deploying agentic workflows
An ADK returns agent has an evalset in which each case expects lookup_order, check_return_window and create_return_label, in that order. A new release adds a get_loyalty_tier call after lookup_order, but only for loyalty customers, and those cases now score 0.0 on tool_trajectory_avg_score at a threshold of 1.0. Which change accepts the extra call and still fails a case when an expected call is missing or out of order?
EXACT is the default match type and rejects any extra call. IN_ORDER requires every expected call in the same order and allows other calls in between. ANY_ORDER would also accept the label being created before the return window check, a lower threshold lets a multi-turn case pass with a skipped call, and an LLM judge gives up the deterministic check.
Section: Developing custom agents
A travel company's ADK agent appends what it learns about each traveler, such as seat choice and dietary needs, to a list under a user: key in session state. Travelers who changed a preference now hear both the old and the new option, the lists keep growing, and the privacy team wants remembered details to expire after 18 months. Which change addresses all three problems with the least custom code?
Memory Bank keeps only information that matches its memory topics, consolidates new information with existing memories in the same scope by updating or deleting what it contradicts, and applies a TTL automatically. Similarity search would return both preferences, app: state is shared by every user, and a summarizing callback is custom code that never expires anything.
Full PAA set: 1 timed exams, every answer explained, $19.99 one-time.
See the practice testsA six-week plan
For engineers who already build on Google Cloud, plan on 60 to 80 hours over six to eight weeks. The schedule below fits six weeks at 10 to 14 hours a week. Build as you go: the hands-on labs after the exam reward the same habits.
Week 1: baseline and the platform map. Take a practice exam cold and keep your score by section. Then read the exam guide and Google's name changes page, and deploy one ADK agent to Agent Runtime with Agent Identity, so every later topic has a place to land.
Week 2: custom agents. Workflow agents and graph workflows, output keys and state prefixes, callbacks and plugins, sessions, Memory Bank and the memory tools. Break things on purpose, such as two parallel agents writing the same key, and watch what happens.
Week 3: knowledge and orchestration. RAG Engine modes, parsers and rerankers, Agent Retrieval with hybrid search, Google Cloud MCP servers and their IAM conditions, Agent Registry, and A2A between agents owned by different teams.
Week 4: low-code and coding agents. CX Agent Studio flows, handoff rules, guardrails and evaluation, Workflow Builder, Gemini Enterprise connectors and sync schedules, then Antigravity rules, skills, hooks, subagents and plugins, the Agents CLI, Claude Code on Google Cloud and GKE Agent Sandbox.
Week 5: evaluation, deployment and security. ADK evaluation criteria, online monitors and quality alerts, Agent Runtime scaling and telemetry, then Agent Gateway, IAM Unified Access Policies, Principal Access Boundary, auth manager, Model Armor and Semantic Governance.
Week 6: rehearsal. Retake the practice exam under time and review every explanation, including the questions you got right. Write a one-line rule for each kind of question that keeps costing marks, and reread the rules before exam day.
Pacing
The generally available exam runs 2 hours. Practice at a pace that leaves time to reread configs and log lines, since a single field often decides the answer.
Mistakes that cost marks
Studying from Vertex AI era material. Older guides describe Agent Engine, Vertex AI Search and Vector Search 2.0 under old names, and they miss the governance layer entirely: Agent Identity, Agent Gateway, Agent Registry and Semantic Governance are new.
Answering with an instruction where the question wants a control. If a scenario says an agent must never take an action, look for a policy, a gateway rule, a hook or a confirmation step. A sentence in a system prompt is a request the model can ignore.
Treating similar controls as interchangeable. A content filter cannot enforce a payment limit, an IAM condition cannot read tool arguments, and a boundary policy cannot tell a small payment from a large one. Know what each control can see.
Forgetting defaults. Many questions turn on a default: an exact match for tool trajectories, a concurrency setting tuned for synchronous code, or a sandbox with no network access. Learn the defaults of the products you study.
Frequently asked questions
Where to start
Read Google's exam guide first, since its sections and in-scope products are the scope of the exam. Then take the free Professional Agentic Architect practice questions without preparing, so the result tells you where to begin. When you are ready to rehearse, the timed Professional Agentic Architect practice exam covers all five sections in Google's proportions, with an explanation for every option.
Sources: