An Agentic Architect Study Plan — Official Samples, English Docs and AI in the Three Weeks Before Exams Open

As of October 2026, GA registration for Google Cloud’s Professional Agentic Architect opens on November 2 and delivery is announced to start in mid-November — about three weeks after this article is published. The exam guide exists only in English, the scope lists freshly renamed and brand-new products, and passing the multiple-choice exam is followed by hands-on labs. An Agentic Architect study plan has to be worked out backwards from those conditions.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article covers where to find questions and study material (official sample questions, the Google Skills learning path, the partner learning path and the community) and how to treat each; how to prepare from English official materials, including for non-native readers; what to check before registering; how to split the three weeks before delivery opens; and how to track what may change at GA. It also decides which week to use the practice from the six articles published October 19–24.

To be clear, we have not taken this exam, did not take the beta and did not submit the official sample-question form. What follows is based on Google and Pearson official pages checked on October 12, 2026 and on the practice we ran ourselves.

Four traits that shape an Agentic Architect study plan

A study plan follows from the nature of the certification. From the exam guide and certification page, this one has four traits:

• Many new products. The exam guide lists 28 in-scope tools, and for Agent Runtime (formerly Agent Engine) and Agent Search (formerly Vertex AI Search) it adds the former names itself
• Hands-on labs. Passing the multiple-choice exam leads to labs on Google Skills, and you are certified only after passing both
• Design judgment. The certification page says the multiple-choice part assesses conceptual knowledge, system design choices and architectural standards. The largest of five domains is developing custom agents (about 33%), and with evaluation and deployment (about 22%) they make up about 55%
• English materials. The exam guide is English only, and the GA language has not been announced as of October 2026

These traits set three directions. First, do not trust AI knowledge as is; give it official materials before asking, because product names and status move month by month. Second, run locally what can be run locally — but local practice does not replace the labs. Third, practice reading in English: use localized resources as support, and check conditions and numbers in the English version.

When we built a study plan for the JDLA G-Test, we also listed the certification’s traits first and fitted learning methods to them (What the JDLA G-Test Proves). The procedure is the same; the methods change. Here, rather than memorizing more, spend time reading changing product names correctly and reading design judgments in English.

The exam as of October 2026, and what is still undecided

Before planning, separate what is decided from what is not. The table shows the GA conditions stated on the certification page and FAQ as of October 2026.

The undecided question count and language matter most for the plan. Among other Professional certifications, Cloud Architect, Data Engineer, Machine Learning Engineer and Cloud Developer can be taken in English and Japanese, but there is no basis for assuming other languages for Agentic Architect, nor for assuming English only. The way to be ready either way is to be able to read in English.

Also, passing the multiple-choice exam does not certify you. The labs are open only to those who pass it and are separate from the labs in the Google Skills learning path. The official materials do not say whether the labs need a Google Cloud account or extra cost, and the FAQ’s GA fee sentence says nothing about tax, whereas the beta’s $120 and Professional Cloud Architect’s $200 carry “plus tax where applicable”.

The recommended experience is three or more years of building, testing, deploying and managing cloud solutions, at least one of them building agents on Google Cloud. It is not an eligibility rule, but three weeks cannot create that experience; what three weeks can do is reconnect the experience you have to the exam guide’s language. The flow from registration to certification is in What the Google Cloud Professional Agentic Architect Certifies, and the basics of Google Cloud certification levels and registration in Google Cloud Certification Guide 2026: The Next Cloud After AWS.

Where to find questions and material — official samples, learning paths, partner training and the community

Apart from the exam guide, the certification page’s Quick links point to four sources (lower on the page it also promotes an agent-development skill badge and a certification-prep webinar). Use them by role, not by volume. Google has stated that it does not publish exam-prep material aimed at passing without building the skills, and that hands-on experience is essential to preparation; choose sources in that spirit.

Official sample questions — learn the format from 10 questions

The sample questions are a Google Form opened from “Review sample questions” on the certification page, titled Professional Agentic Architect Sample Questions: 10 single-answer questions with four options each, one per page, 12 pages in all. The page before the questions has six required fields (first name, last name, email, affiliation, relationship to Google, and whether to receive emails). A Google sign-in is needed only to save progress.

The form’s description states their role: to familiarize you with the format of exam questions and example content that may be covered; they do not represent the range of topics or level of difficulty, and you should not use your results to predict your exam result. So do not use your score on 10 questions to forecast pass or fail.

Submitting requires those fields, so you send your name and email to Google. The post-submission message says results can be viewed with “View Accuracy”, but we did not submit, so we have not confirmed whether correct answers or explanations are shown. The Japanese-language certification page labels this link as a “practice exam”, but it is the 10-question sample; we call it the official sample questions.

Google Skills learning path — not mandatory, a starting point

“Train for the exam” on the certification page opens the learning path “Professional Agentic Architect Certification” on Google Skills (formerly Google Cloud Skills Boost). As of October 2026 it has 12 activities, all courses, one of which (Engineer AI Agents with Agent Development Kit (ADK)) is a skill-badge course; listed durations total a little over 40 hours. Courses contain labs and quizzes — the first, in English, has four labs and seven quizzes.

The official materials state two things about the path. It is a preparation resource and not mandatory (FAQ). And it is a starting point: candidates are expected to have hands-on experience and a full understanding of every item in the exam guide (path description). The FAQ also distinguishes the path’s labs from the exam labs that only those who pass the multiple-choice exam take. Finishing the path neither covers the whole guide nor completes the certification labs.

Cost depends on how you use it. Browsing the catalog needs no sign-in, but starting the path or a course does, and most labs cannot be opened without credits or a subscription. The free Starter tier gives 35 credits a month with limited labs; the paid Pro tier is $29 a month (seven-day trial for new subscribers) with unlimited labs. One credit is roughly one US dollar, the number needed depends on lab difficulty, and credits cannot be bought on their own. How many credits each lab in this path needs is visible only after signing in, and we did not check, so confirm in your own account whether the free tier is enough.

Partner learning path — same content as the public one

The certification page also lists “Partner training for the exam”, a learning path on Google Skills for Partners. As of October 2026 the names and durations of its 12 activities all matched the public path, and the FAQ recommends it to partners. Judging by names and durations, either path leads to the same courses, though the partner path’s description also recommends the Certified Partner Specialist program, which we did not look into.

The community — not a certification-only forum

“Join the learning community” leads to the Google Cloud category of the Google Developer forums, described as a community for people building, connecting, deploying and running on Google Cloud. It has 11 subcategories (Build with AI and others), none dedicated to certification. Google recommends it to certified people as the best place to hear about beta exams. For GA preparation, use it for announcements and product discussion, not for asking about exam content.

What does not count as a source

The exam terms define exam content as Google’s confidential information (§3), prohibit sharing it, including what you remember, on posts, forums or blogs (§5.a.vi), and prohibit using or viewing unauthorized exam content shared by others (§5.a.vii). Candidate write-ups that touch on actual questions are neither sources nor references. This article does not reproduce the sample questions or build practice questions from them.

Treat the English official materials as authoritative — localized pages and courses lag

Localized official pages are a convenient entry point: the certification page, the FAQ and Google Cloud certification help are available in other languages, including Japanese. Even so, use the English versions as the basis for conditions and numbers, because real differences exist.

First, the Japanese help and FAQ carry a note that AI translation may contain errors. Second, translation has changed meaning: the beta entry on the English certification page says “~80 multiple choice questions” (about 80), while the Japanese page says the equivalent of “up to 80”. “About” and “up to” are not the same, and the sample-question link is translated as “practice exam”, as noted above.

Courses differ too. The first course in the path, Understand Google Cloud Agents, lists Vertex AI Agent Builder, Gemini Enterprise, conversational agents and the Agent Development Kit in its Japanese description, while the English description lists Gemini Enterprise app, Conversational Agents, Customer Experience (CX) Studio, Gemini Enterprise Agent Platform and the Agent Development Kit, with no Vertex AI Agent Builder. In the outline, where the English version has two chapters (CX Agent Studio and Gemini Enterprise Agent Platform), the Japanese version has a single “Vertex AI Agent Builder” chapter — it appears to be an older version left in place.

The last course is similar: the English title is Deploy Multi-Agent Systems with Gemini Enterprise Agent Platform, while the Japanese title still uses Agent Engine, the former name of what the exam guide calls Agent Runtime. Two of the 12 courses keep English titles even in Japanese display, and some durations differ by display language (Manage Agent Memory and State shows 45 minutes in English and 1 hour in Japanese).

Whether video and lab content is localized can be seen only after signing in, and we did not check, so we cannot say the whole path can be taken in another language. The exam guide is English only, and even the Japanese certification page links to the English PDF. If you study from localized material, keep a table at hand for mapping old names to the exam guide’s names — the table in the next section serves that purpose.

Give the AI official material before asking — build and use a name table

If you use AI to read English materials, order matters. Do not ask the AI first and then check against official materials; give it the official material first, then ask. In Agentic Architect Exam Scope: What AI Knows About 28 Tools, we had claude -p (recorded model ID claude-sonnet-5-5) explain the 28 in-scope tools with no materials and with the official page attached, and judged the answers against official docs retrieved October 12, 2026, one run per condition.

In that run, definitions were correct for 18 of 28 items without materials and all 28 with the official page. Of the 10 items with former names, current names were given for 4 without materials (old names for 3) and for all 10 with the page. The model’s knowledge cutoff is June 2026 (both the reliable knowledge cutoff and training data cutoff in Anthropic’s model list), and the Agent Platform rename (April 22, 2026) came before it; even so, without materials it answered with the names Vertex AI Model Garden and Vertex AI RAG Engine.

Release status (GA or preview) had another trap. With the official page attached, 22 of 24 scored items came back Unknown (10 without materials). Google Cloud docs show a Pre-GA notice at the top of preview pages but nothing equivalent at the top of GA pages (some pages, like RAG Engine’s, include a table of GA/preview status by region; that item varies by region, so it was excluded from status scoring). If the AI must rely only on the material, anything not labeled GA comes back Unknown. And if the Preview notice at the top is dropped when passing page text to the AI, preview products become indistinguishable too — our first run with pages attached lost that notice in extraction, so we fixed it and reran.

From that check we built a name table. Product names stay in English (Google’s localized docs do not translate them either), alongside the former name and a paraphrase of the official definition. Seven entries, including renames the exam guide does not annotate:

Two more points beyond the table. Agent Platform is short for Gemini Enterprise Agent Platform; Google describes it as the evolution of Vertex AI, renamed on April 22, 2026 — the official words are “evolution” and “now part of”, not discontinued. And Gemini Enterprise, the search-and-assistant product for business users, is a different product from Gemini Enterprise Agent Platform, the development platform; the names are similar and easy to confuse when reading English.

Old names also survive in official materials: Agent Runtime’s API resource is still ReasoningEngine, Agent Search’s API is the Discovery Engine API, Sensitive Data Protection’s is the DLP API, and “Agent Engine” still appears in release notes after the rename. An AI that answers with an old name is not necessarily wrong; the table exists so that either name leads you to the same product.

A request to the AI might look like this:

Using only the body of the attached official page, answer the following about the product below.
(1) A one-sentence definition (2) Its current official name and former names (3) Its release status (GA or preview)
For anything not in the material, write “not stated in the material” and do not fill gaps with guesses.
The Preview notice at the top of the page is included in the body.
Product: Agent Gateway

Handing the model the right context first, and telling it exactly how to treat what the context does not cover, is the core of prompting with documents. A practical book on prompt design for LLMs goes deeper into this.

Prompt Engineering for LLMs

USD 54.51 on Amazon.com (as of 2026/09/22)

Amazon

ポチップ

Practice reading questions in English, with AI

Since the GA language is undecided, be ready to read English questions as English. Google Cloud certification states plainly that time extensions do not apply because English is a second language — not merely unmentioned, but explicitly excluded. If you sit the exam in English, you must find your reading speed within the two hours.

Exam conditions shape practice too. In online exams you cannot read questions aloud and cannot use books, paper, pens or phones; notes go on a digital whiteboard. The exam is not open-book, and there are no breaks. So practice reading English continuously, silently and without paper. Run it in this order:

1. Read the exam guide in English. Read the five domain headings (Developing custom agents and so on) and subitems without translating, marking sentences you cannot parse
2. For marked sentences only, have the AI break down the structure — not translate it, but separate what is asked, the conditions and what is excluded
3. Look product names up in the name table; when asking the AI, attach the table and the relevant official page
4. Do the official sample questions once without AI. Time yourself, and for each question that made you hesitate, note why (word meaning, reading the conditions, product knowledge)
5. Reread those spots with the AI using the same request as step 2. Use the sample text only within your own practice and do not share rewritten questions

An example request for steps 2 and 5:

Do not translate the following English passage.
Split it into (1) what is being asked (2) the conditions (3) what is excluded, quoting the English phrases as they are.
Use the product names as written in the attached name table, and mark any term not in the table as “not in table”.

There is a reason not to translate. Reading a translation can feel like understanding, but if you sit the exam in English, English is what you will see (the beta was English only, and the GA language is unpublished as of October 2026). Having the AI quote English phrases verbatim helps you spot which words set the conditions. If you hesitated over a word’s meaning, add it to the name table; if over reading the conditions, find sentences of the same shape in the exam guide and reread them.

Turn the words you hesitated over and misread product names into cards with the method in Log Your Mistakes, Let AI Sort Them, Push Them to Anki, and you will see them repeatedly over the three weeks. Put the English product name on the front and the former name and definition on the back, and the name table becomes your repetition material.

What to check before registering — name in Roman letters, ID and accommodations

Registration opens November 2 in CM Connect; you do not create a separate Pearson account. During registration you choose the delivery method (test center or online) and the exam language. The procedure is short, but three checks cannot be undone later.

First, your name. The legal name in CM Connect must match a government-issued photo ID, and for Pearson scheduling it must be entered in Roman characters. A mismatch means you cannot test and the fee is not refunded. The ID needs your name, photo and signature. The official materials do not say whether an ID with your name only in a non-Latin script is accepted, so decide before registering which ID you will match and register the Roman spelling based on it. For a test center, bring two forms of ID.

Second, accommodations. Individual accommodations for disabilities and similar needs are requested in CM Connect before booking, and you cannot book until approved. Review takes 5–7 business days, and arrangements can take at least 15 days for online exams and up to 30 days at test centers, depending on the request. If you want to test right after delivery opens, apply early. As noted, English as a second language does not qualify for extra time.

Third, age and rescheduling. Candidates must be 18 or older (terms §2.a), and rescheduling or cancelling within 24 hours of the exam incurs a fee. Know the retake policy too: for Google Cloud Associate and Professional certifications you can take an exam up to four times in two years, with waits growing to 14, 60 and 365 days after each failure, and you pay every time. Changing language or delivery method still counts as an attempt, while beta attempts do not. Results show only pass or fail with no score, but a failed GA attempt shows per-section results (the beta does not).

Set up the online exam environment first

Online exams run on Pearson’s OnVUE, and the fee is the same either way. If you plan to test at home, check the environment requirements before registering — finding out on exam day that you do not meet them cannot be fixed.

The easy-to-miss items are network and location. Pearson’s Google Cloud page lists corporate networks and VPNs as unusable and rules out public places such as offices, libraries and cafes (Google’s registration help says you can test from home or work, but Pearson’s rule is stricter). If you lack a qualifying PC and connection at home, a test center is the safer choice.

Google’s help and Pearson’s page describe check-in timing differently, so follow your booking confirmation and start early. They also differ on drinks: Google allows none during the exam, while Pearson allows a drink in a container with no text or labels on the desk; the table follows Google’s stricter rule. Online-exam support is available in English, Japanese and Spanish, but that is the support language, not the exam language.

Splitting the three weeks before delivery opens

Now the plan. The three weeks are an editorial division of the time from this article to mid-November, not a study period Google recommends. You choose your exam date when you register; if three weeks is not enough, book a later date. The idea is simple: put the heaviest domains, §3 (about 33%) and §4 (about 22%), in the hands-on second week.

Week 1 — align scope, terms and format

The first week puts the scope in your head in English. Save the exam guide PDF with the retrieval date and read the five domain headings (about 13, 17, 33, 22 and 15%) and subitems in English. Then build the name table, starting with the renamed items among the 28 and similar names that are easy to mix up (Gemini Enterprise versus Gemini Enterprise Agent Platform, for example).

Do the official sample questions once this week, to learn the format and the kinds of things that make you hesitate — not for the score. Finish the pre-registration checks (ID and Roman-letter name, online environment, accommodation requests) this week too, so you can register on November 2 without hesitation. Browse the learning-path catalog and pick the courses for weeks two and three.

Week 2 — get hands-on (§3, §4, §2)

In the week registration opens, register and set your date, then spend the rest of the time building. Custom agents, about 33% of the exam, can be practiced with the steps in Run ADK Agents Locally with Ollama. We connected ADK for Python v2.11.0 to a local Ollama model (gemma4:12B, written gemma4:12b on Ollama’s model page) and wrote and ran an agent with tools plus sequential and parallel workflows. In that run, the tool agent called tools in the instructed order on all five questions in both recorded runs.

Locally you get only the mechanics, though. Memory Bank, Agent Identity, Agent Registry and search with Vector Search or Agent Retrieval are Google Cloud services that cannot run locally, so read about them in the official docs. Local practice does not replace the labs.

For evaluation and deployment (about 22%), ADK Evaluation Locally: What an Evalset Catches and Misses ran ADK evaluation locally. A deliberately broken agent that reserved without checking was caught by every matching method, while for an agent with an extra tool call, only the default matching (exact match of names, arguments, order and count) caught it.

Response evaluation needs care too. Computing the metric that counts overlap between a response and the reference (ROUGE-1 F-score, counted per character for Japanese) on fixed sentence pairs, the opposite-meaning sentence (0.772) scored higher than the paraphrase (0.464). Knowing in numbers what an evaluation metric does and does not look at helps when reading evaluation designs in English.

For coding agents (about 17%), practice with Claude Code Settings Tried Locally. We configured local Claude Code (v2.1.296, Windows 11, normal login) with an MCP server, a skill, rules, hooks, a subagent and a plugin, once each, and in that run the reservation tool call was stopped by a hook. Remember that §2.2 is an Antigravity item; practicing it with Claude Code is our substitution.

Week 3 — train design judgment and finish (§5, §1)

The last week practices security and governance (about 15%) and low-code tools (about 13%) as design judgment. Agent Governance Design: AI Drafts Scored with a 15-Row Rubric had claude -p solve our own design problem (not an exam question) and scored it with a 15-row rubric built from official best practices.

In that run, the design given only the problem met 7 of 15 rows and the one given §5’s items met 10 of 15; the rows both missed included features that went GA after the model’s knowledge cutoff. Here too, the design given official material (the §5 items) met more rows.

Spend this week on repetition and rechecks as well: run the cards of words you hesitated over since week one. Do the sample questions again, timed, if you like, without reading pass or fail into the score. Do one session of reading English official material for two hours straight with no break, as in the exam. Finally, check the seven places in the next section for changes at registration and delivery.

If you want to take the week-two practice beyond a local ADK agent, a practical book that builds agents with tools and grows them into multi-agent systems is a good next step.

AI Agents in Action, Second Edition

USD 47.48 on Amazon.com (as of 2026/09/22)

Amazon

ポチップ

Tracking what may change at GA

The conditions in this article are as of October 2026 and may change around registration and delivery. Terms §1.e lets Google update exam content at any time to keep pace with Google Cloud technology, and the certification list carries a banner saying exams are being revised to reflect product updates announced at Google Cloud Next ’26 (a notice for all exams, not only Agentic Architect). What may change, and where to look:

The exam guide takes the most effort: its PDF has no revision date, so dates tell you nothing. Save a copy with the retrieval date and compare the text with newly retrieved versions. Giving both versions to an AI to list differences saves effort, but always confirm each listed difference in the PDF itself.

Page wording can linger or change without notice. On October 12, 2026 the learning path description still said, as written before the exam guide was released, to watch for registration details and the exam guide in early September. When reopened on October 13 that sentence was gone, replaced by one saying the path will be updated for GA in November. Not every sentence on an official page is current, and a sentence you read before may not remain. We will review this article against the official pages after registration and delivery open.

What this preparation cannot tell you

• Whether it is enough to pass. We have not taken the exam, and the GA question count and language are unpublished. We cannot say this plan will get you certified
• The correct answers and explanations of the sample questions. We did not submit the form and do not know whether answers or explanations appear afterwards
• Whether the learning path can be finished for free. We did not check credits per lab
• Who can use the partner learning path. We found no official page stating eligibility
• The language of localized course content. Videos and labs open only after signing in, and we did not check
• Generalizations from our practice. Each article’s runs were one or two per condition, not general properties of AI or tools
• The content and cost of the labs. They are for those who pass, and the official materials do not say whether a Google Cloud account or extra cost is needed

Summary — build your Agentic Architect study plan around giving AI the official material first

Build the plan from the certification’s four traits: many new products, hands-on labs, design judgment and English materials. Use the four sources by role: the official sample questions (10 questions, to learn the format), the Google Skills learning path (12 activities, a starting point, not mandatory), the identical partner path, and the community for announcements.

Treat the English version as authoritative, and use localized pages and courses as support while watching for old names and translation differences. Give the AI official material before asking: in our run, definitions were correct for 18 of 28 items without materials and all 28 with the official page. Read questions in English, and have the AI break down the structure only of sentences that make you hesitate.

Before registering, check your ID and Roman-letter name, your online-exam environment and any accommodation request. Spend the three weeks on scope and terms, hands-on practice, then design judgment and finishing, and end by checking the seven places for changes. Much is still undecided as of October 2026, so keep the plan easy to rearrange around the official pages.

Sources

• Professional Agentic Architect certification page
• Professional Agentic Architect Certification FAQs
• Professional Agentic Architect exam guide (PDF)
• Professional Agentic Architect Sample Questions
• Google Cloud certifications
• Google Skills learning path: Professional Agentic Architect Certification
• Google Skills for Partners learning path
• Google Skills subscriptions
• Google Skills Help: Subscriptions
• Google Developer forums: Google Cloud
• Registration and Scheduling
• Online Proctored Exams
• Testing requirements for proctored exams
• Acceptable forms of identification
• Requesting an accommodation
• Retake Policy
• Rescheduling or canceling a scheduled exam
• Localized language exams
• Google Cloud Exam Terms and Conditions
• Pearson VUE: Online testing for Google Cloud
• Professional Cloud Architect certification page
• Gemini Enterprise Agent Platform name changes
• Anthropic: Models overview

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.