Beta window · 8–30 September 2026

The exam scores
your reasoning.
So does this.

Scenario questions for the Google Cloud Professional Agentic Architect, weighted exactly as the official exam guide weights its five sections. Every option is explained — including the wrong one you were about to pick. Ten are free on this page.

Progress saves in this browser · no account · no sign-in

What you get

10Questions free, right nowNo sign-up. The full set is 50, unlocked with a key for €10.
4Explanations per questionEvery option gets a reason, not just the correct one.
5Sections, correctly weightedThe free ten follow the same 13 / 17 / 33 / 22 / 15 split the exam guide publishes.
3 modesTimed mock, study, section drillSit all fifty against the clock, or work one section at a time.
BrowserNothing to installOpens like any web page. Your progress saves on your own device.

How it teaches

Not an answer key. A reason for every option.

Most banks tell you the letter. The letter is worthless on a question you have not seen. Here is one, exactly as it appears.

Section 3 · Developing custom agents Architect-level

Policy states an agent may never issue a refund above $500 without human approval. Which implementation actually enforces it?

AA strongly worded instruction stating the $500 limit, repeated at the start and end of the system prompt.

BA deterministic check in a pre-tool callback that inspects the refund amount and blocks or routes to approval before the tool executes.

CA post-response filter that scans the agent's reply for refund amounts over $500.

DA model fine-tuned on examples that always escalate refunds above $500.

APrompt instructions are probabilistic and are the first thing an adversarial user attacks. A policy that fails on a clever phrasing is not enforced.

BThe check runs in code, before the tool call reaches the payment system, and holds regardless of what the model was persuaded to attempt.

CBy the time the reply is scanned, the refund tool has already run and the money has moved.

DFine-tuning shifts the distribution of behaviour. It does not create a hard bound, and it fails silently on unfamiliar inputs.

Rule to carry forward Hard business limits belong in deterministic code on the execution path, not in prompts or model weights. The model proposes; code disposes.

Coverage

Weighted the way the exam is weighted.

Section 3 is a third of the paper. An evenly split practice set trains you for an exam that does not exist — so the question counts here follow the published blueprint, in the free ten and in the full fifty alike.

Scope

The 28 tools the guide names.

If a service is not on this list, it is not what the question is testing. Two were renamed for this exam — answers written against the old names read as wrong.

The full set

Ten here. Forty more for €10.

The ten questions on this page are the whole product in miniature — same scenarios, same four explanations, same rules to carry forward. The other forty complete the blueprint coverage. One payment, no subscription, no account.

€10one payment · VAT included

All 50 questions, weighted 13 / 17 / 33 / 22 / 15
Four explanations on every question
Timed mock, study mode and section drills
Your key works on any browser you use

Buy the full set

Your key arrives by email the moment you pay.

Already have a key?

Paste it below. It is checked once and then remembered on this device.

Start

10 questions ready now. Nothing to install.

Sit the set against the clock, or take one section at a time. Every answer you give is explained before you move on, and your progress stays on your own device.

No account · no sign-in · opens in this browser

New certification, new set

This is the first of these. If you want to know when the next one lands — another certification, or more questions for this one — leave your address. Nothing else, and an unsubscribe link on every message.

See how this address is handled.