Professional Agentic Architect

Last checked November 10, 2026 against the official Google Cloud exam page.

1. Question 1Developing custom agents

   In the Agent Development Kit (ADK), which workflow agent type runs its sub-agents one after another in a fixed order?

   1. ASequentialAgent, which runs its sub-agents one after another in the order they are listed
   2. BParallelAgent, which runs its sub-agents concurrently and then waits for all of them to finish
   3. CLoopAgent, which repeats its sub-agents until a stop condition or an iteration limit is reached
   4. DLlmAgent, which lets a model reason and choose tools for a single task on its own
   Show answer and explanationHide answer

   Correct answer: A

   SequentialAgent executes its sub-agents in the listed order, which matches the requirement. ParallelAgent runs them concurrently, LoopAgent repeats them, and LlmAgent is a model-driven agent rather than an orchestration pattern.

2. Question 2Evaluating and deploying agentic workflows

   A team's agent scores well on its evaluation set but fails in production on long, multi-turn conversations. Which change to the test set best addresses this gap?

   1. AAdd more single-turn questions with short answers, since they are easier to score consistently across runs
   2. BAdd multi-turn conversations and edge cases that resemble real user sessions, including the failing patterns
   3. CRemove the production cases that the agent fails, so that the pass rate reflects the more typical everyday use
   4. DLower the pass threshold until the evaluation results more closely match what the team sees in production dashboards
   Show answer and explanationHide answer

   Correct answer: B

   The evaluation set must cover the situations where the agent fails, so multi-turn and edge-case conversations drawn from real sessions reveal the gap. Adding easy single-turn items, removing failing cases or lowering the bar would make the score look better without measuring the problem.

3. Question 3Using coding agents for application development

   A developer lets a coding agent refactor a service from a Cloud Workstation. The agent's environment currently holds a long-lived key for the production database. Which change best reduces the risk?

   1. AKeep the production key in the environment, since the agent needs it to test database changes end to end against live data
   2. BAdd a note to the service's README asking developers not to run the agent's migration scripts against production
   3. CRun the agent in an isolated sandbox with short-lived, least-privilege credentials that cannot reach production data
   4. DTurn off code completion in the IDE for the duration of the refactor so the agent's edits are easier to review
   Show answer and explanationHide answer

   Correct answer: C

   An isolated sandbox with short-lived, least-privilege credentials limits what a faulty or manipulated agent can reach. Keeping the production key, relying on a README note, or changing IDE settings does not contain the risk.

4. Question 4Securing and governing agentic workflows

   An agent helps employees check their own leave balances by calling an HR system's API. Which design follows best practice for authorizing the agent's tool calls?

   1. AAsk employees to type their HR password into the chat so that the agent can sign in as them for each request it makes
   2. BStore one shared administrator service-account key in the agent and use it for every employee's request to the HR system
   3. CHave each employee authorize the agent via OAuth 2.0, request only the read scope it needs, and keep tokens server-side
   4. DGive the agent an unrestricted token for the HR system so that it never fails with a permission error during a conversation
   Show answer and explanationHide answer

   Correct answer: C

   User-delegated OAuth 2.0 with least-privilege scopes means the agent can only read what each employee is allowed to read. A shared admin key, pasted passwords or an unrestricted token all remove per-user limits and create a larger risk if the agent is misused.

5. Question 5Building agents using low-code tools

   A retail company is building a customer-support agent in CX Agent Studio. The conversation must follow a strict order: verify identity, collect the order number, then check shipping status, and the verification step must never be skipped. Which design best meets this requirement?

   1. APlace every instruction in one long system prompt and let the model decide the order of the steps on each turn
   2. BModel each stage as a page with its own instructions, and add transition routes that advance only when the verification condition is met
   3. CCreate a separate agent for each possible customer question and have a support engineer route each conversation to the right one by hand
   4. DKeep the full chat history in a database and have the model re-read it on every turn to infer which step the customer is currently on
   Show answer and explanationHide answer

   Correct answer: B

   Pages with transition routes enforce a fixed order because a route fires only when its condition holds, so verification cannot be skipped. A single system prompt cannot guarantee the sequence, a separate agent per question is impractical, and re-reading history leaves the current step to guesswork rather than a defined state.