CertifyDeckCargando contenido…
Professional Agentic Architect · Sample questions
10 real questions from the CertifyDeck Professional Agentic Architect bank, spread across the domains of the official Google Cloud Certified Professional Agentic Architect skills outline, with the correct answer and a full explanation of why the other options are wrong. They are original questions written from the public syllabus, not exam dumps, and they are the same style you will find in the practice tests and mock exams.
Correct answer: A. Run cm fix with the --context flag and describe the constraint to use the parameterized query helper.
The -c or --context flag of cm fix passes architectural constraints or coding guidelines to the agent, such as using a specific safe parameterized database helper, so the generated patch follows them without manual edits. --no-cache only forces a fresh patch with the same approach, and disabling the sandbox or the command policy weakens host protection without telling the agent which helper to use.
Correct answer: C. Set authProviderType to google_credentials so that each developer's Application Default Credentials are attached.
Setting authProviderType to google_credentials makes Antigravity attach Google Application Default Credentials, which each developer sets up locally with gcloud, so no token is stored in the file and calls run as the developer. Moving the bearer token to the global file still leaves a static, expiring token in configuration. A service account key in the repository is a long-lived credential that is not the developer's identity, and a client secret in the shared workspace file would be committed.
Correct answer: D. Register the auth provider oauthcallback URL as a redirect URI in the GitLab application.
When you configure the OAuth client for a 3-legged auth provider, you must register the auth provider's dedicated callback redirect URI, which ends in oauthcallback, on the third-party OAuth application. The continue URI is where your own frontend receives the user after consent, so pointing it at GitLab does not fix the redirect check. A 2-legged provider would remove the developer's delegated authority that the scenario needs.
Correct answer: A. Grant the Agent Identity User role on the auth provider resource to the agent principal.
The agent authenticates to auth manager with its own SPIFFE identity, and access to each auth provider is governed by IAM; the documented grant is the Agent Identity User role on the auth provider resource for the agent's principal. Granting it to a project-wide principal set would let every agent use the provider, and the Admin role adds management permissions the agent does not need. The service agent is not the identity that retrieves the credentials.
Correct answer: B. Add a vertex_embedding_config with a model and a text template to the vector field in the Collection's vector schema.
In Agent Retrieval, a vector field whose schema includes vertex_embedding_config is populated automatically, using the configured model and text template, whenever a Data Object is created without a value for that field. A sparse-only field matches keywords, not meaning, which fails the requirement. Relaxed schema validation controls how strictly data fields are checked and does not generate embeddings at search time.
Correct answer: C. Enable Claude from its Model Garden model card and pass the Claude model string to the LlmAgent to call the managed API.
Claude models on Agent Platform are fully managed, serverless APIs that support Agent Platform request-response logging, and ADK's registry routes a Claude model string to its Claude wrapper on Agent Platform. LiteLlm with an Anthropic API key calls the Anthropic API directly, so requests bypass Agent Platform, while a dedicated endpoint or a GKE cluster means provisioning and managing serving infrastructure.
Correct answer: B. Create a Cloud Monitoring alerting policy on the scores exported by the monitor, with an email notification channel.
Online monitors export numeric evaluation scores to Cloud Monitoring, where alerting policies on those metrics notify notification channels such as email, and each incident carries labels for the associated monitor. A test_config.json threshold only affects ADK eval runs in development, an hourly CLI run adds operational effort and does not score live traffic, and Model Armor screens prompts and responses for safety rather than measuring task success.
Correct answer: C. Enable Agent Anomaly Detection for the agents so that their logs and traces are analyzed out of band.
Agent Anomaly Detection analyzes agent logs and traces asynchronously, so it adds no latency, and its detectors cover cascading failures such as infinite loops and oscillating retries as well as resource exhaustion; findings are surfaced in Security Command Center. Semantic governance and Model Armor are inline checks on the request path and target intent or content, not runaway execution. max_iterations requires code changes and produces no security findings.
Correct answer: D. Reorder each example so that the step-by-step reasoning comes before the final JSON verdict.
Google's prompt health checklist flags incorrect chain-of-thought order: examples should not show the model producing its final structured answer before it has completed its step-by-step reasoning. Putting the reasoning first lets the verdict follow from it, and separators or a defined output format can still keep the reasoning apart from the answer. Removing the reasoning gives up the benefit of explaining it, while more examples or moving them into system instructions keeps the faulty order.
Correct answer: A. Add the readOnlyHint annotation to the read-only tool definitions, and then reload the custom actions.
By default every action imported from a custom MCP server requires user confirmation because Gemini Enterprise assumes it may mutate data; adding readOnlyHint to non-destructive tools makes them bypass confirmation, and reloading custom actions applies the change. destructiveHint does the opposite and keeps confirmation, which is right only for the tools that modify data. Confirmation is driven by tool annotations, not by an IAM role, and turning off all actions removes the capability instead of streamlining it.
Answer each question before opening the explanation, and write down why you discarded the other options. On Professional Agentic Architect, most mistakes come from choosing an option that is technically possible but does not meet every requirement in the scenario.
These 10 questions are a taste of the format, not a diagnostic. To know which domains you fail, practice by domain for free and take a timed mock exam that mirrors the official weights and duration.
Start for free, no sign-up or card. Every question includes an explanation like the ones above, and the mock exam reproduces the official Professional Agentic Architect format.