Why practice questions matter for this exam
The Google Cloud Professional Agentic Architect exam goes GA in mid-November 2026 (the beta closed September 30). It's a professional-level, scenario-heavy exam: ~2 hours, US$200, multiple choice and multiple select, plus a hands-on lab component — and Google doesn't publish an official passing score. The questions below are built from the official exam guide domains: custom agents with ADK, agent protocols (A2A, MCP), retrieval and grounding, agent identity and governance, runtimes, and evaluation. Work through them under exam conditions: read the business requirement first, eliminate two options, then commit.
Don't sit the real exam until you're consistently passing full-length practice tests comfortably. These 10 are a warm-up — the complete timed practice exams are linked at the end.
Question 1 — Custom agent framework
A retail company wants an agent that monitors inventory across three warehouses, negotiates restocking with supplier agents, and triggers purchase orders through a custom ERP API. The flow requires custom orchestration logic, retries, and stateful multi-step coordination. Which approach should the architect choose?
- A. Gemini Enterprise with a no-code agent designer
- B. A custom agent built with the Agent Development Kit (ADK)
- C. A Gemini API prompt with function calling only
- D. A Cloud Workflow with hardcoded branching
Answer: B. Custom orchestration logic, multi-agent negotiation, retries, and stateful coordination are exactly what ADK is for — you own the code, tools, tests, and deployment. A is wrong because no-code designers suit structured, repeatable flows and can't express custom multi-agent negotiation. C is wrong: bare function calling has no orchestration, state, or eval story. D is wrong: Cloud Workflows orchestrates services, not agentic reasoning loops.
Question 2 — A2A vs MCP
An agent must (1) delegate a research subtask to a specialist analyst agent, then (2) write the findings into Salesforce. Which protocols cover each leg?
- A. MCP for both legs
- B. A2A for leg 1, MCP for leg 2
- C. A2A for both legs
- D. REST for leg 1, A2A for leg 2
Answer: B. Agents talking to agents = A2A (agent-to-agent communication and handoff). Agents calling tools or data = MCP (Salesforce is a tool/data connection via an MCP server). A and C are wrong because they mix up the two protocol scopes; D is wrong because plain REST gives you no agent discovery, task lifecycle, or capability negotiation.
Question 3 — Retrieval approach
An agent must answer employee questions from 2 million internal documents with minimal operational overhead. The team has no vector-database expertise and wants a managed end-to-end pipeline. Which retrieval approach should they choose?
- A. Custom RAG on Vector Search with hand-tuned chunking
- B. RAG Engine (managed RAG pipeline)
- C. Full document text pasted into the prompt on every turn
- D. BigQuery with LIKE queries over document text
Answer: B. Minimal ops + no vector expertise = managed retrieval first; RAG Engine is the end-to-end managed RAG pipeline. A is wrong here because custom chunking/reranking is only worth it when you can name the specific control you need — and you own the pipeline, refresh, and evals. C fails on context limits and cost at 2M documents. D is keyword search, not semantic retrieval, and won't ground answers reliably.
Question 4 — Agent identity and least privilege
A fleet of agents acts on production systems on behalf of different users. The security team requires each agent to have its own identity with permissions bounded to exactly the resources its task needs — no shared service-account keys. Which design meets the requirement?
- A. One shared service account with broad Editor access for all agents
- B. Agent Identity for each agent, bounded by principal access boundary (PAB) policies
- C. API keys embedded in each agent's configuration
- D. OAuth 2.0 tokens with no scoping, rotated yearly
Answer: B. Per-agent identity (Agent Identity) bounded by principal access boundary policies is the least-privilege design for agents. A violates least privilege and creates a single blast radius. C is the classic anti-pattern — long-lived shared secrets. D is wrong because unscoped tokens plus yearly rotation is the opposite of short-lived, least-privilege access.
Question 5 — Runtime selection
A production agent needs managed conversation sessions, long-term memory, and evaluation hooks. Traffic is steady during business hours with no extreme spikes. Which runtime is the best fit?
- A. Cloud Run
- B. Google Kubernetes Engine (GKE)
- C. Agent Runtime
- D. Compute Engine VMs
Answer: C. Managed sessions, Memory Bank, and evaluation hooks are Agent Runtime's native features — match the runtime to the requirements. A is wrong: Cloud Run is great for stateless, spiky, scale-to-zero endpoints, but you'd wire sessions and memory yourself. B is wrong: GKE is for custom networking, GPUs, and full control — overkill here, and you manage the cluster. D is wrong: raw VMs add ops burden with no agent-specific value.
Question 6 — Evaluation strategy
Before launch, the team must prove the agent still handles edge cases after every prompt change. After launch, they must catch quality drift in production. Which evaluation pairing is correct?
- A. Manual spot-checks before launch; no production evaluation
- B. ADK evalset in the dev loop; continuous evaluation pipelines in production
- C. Production A/B tests only, no pre-launch evals
- D. A single golden prompt reviewed once by the team
Answer: B. Evalset (golden data, prompts, edge cases) runs on every change in the dev loop; continuous evaluation pipelines catch metric regressions in production like SLOs. A and D don't scale and can't prove anything about edge cases. C is backwards — you need pre-launch regression before exposing users, and A/B tests alone don't detect drift systematically.
Question 7 — Low-code vs custom build
The business wants an assistant over internal HR documents. The flow is structured (greeting, authenticate, search docs, escalate to a human), speed matters, and the team may want custom logic later. What should the architect recommend?
- A. Custom ADK agent from day one
- B. Gemini Enterprise agent designer for the structured flow, knowing custom logic beyond a state machine needs ADK later
- C. A static FAQ page
- D. Fine-tune a model on the HR documents
Answer: B. Structured, repeatable process + speed matters → low-code first — with eyes open that custom logic beyond the state machine means ADK later. A is wrong: building custom on day one wastes the speed advantage when the flow is structured. C doesn't meet the conversational requirement. D is wrong: fine-tuning is for behavior/style, not for document Q&A — retrieval (RAG) is the right tool for grounding on HR docs.
Question 8 — Guardrails for risky actions
An agent can issue refunds up to $500 without human review. The business wants to keep the automation but add protection against prompt-injection-driven refund fraud. Which control combination is appropriate?
- A. Remove all refund capability from the agent
- B. Model Armor screening on prompts/responses plus human-in-the-loop approval for refunds above a threshold
- C. A longer system prompt asking the agent to "be careful"
- D. Rate-limiting the agent to one refund per day
Answer: B. Screen with Model Armor and gate the high-risk, irreversible action with human-in-the-loop approval — the standard guardrail pattern for agentic systems. A destroys the business value. C is not a control — prompt instructions don't stop injection. D limits volume but doesn't distinguish legitimate refunds from fraud.
Question 9 — Workflow pattern
A research agent must gather data from five independent sources, then synthesize the results, then route the summary to a specialist based on topic. Which workflow pattern fits?
- A. Strictly sequential: source 1 → source 2 → … → source 5 → synthesize
- B. Parallel fan-out for the five sources, then sequential synthesis, then conditional routing (graph)
- C. A single LLM call that does everything
- D. Random-order polling of the sources
Answer: B. Independent work fans out in parallel; dependent steps run sequential; dynamic routing to the specialist is a graph decision. A wastes time serializing independent calls. C will blow context limits and can't reliably coordinate five sources. D is not a pattern — it's chaos.
Question 10 — Cost-efficient model selection
A ticket-classification agent handles 10 million requests per month. The task is narrow, latency-sensitive, and has a strict cost target; data must stay in-region. Which model strategy is best?
- A. The largest available SaaS LLM for every request
- B. A fine-tuned small language model (SLM) for the classification task
- C. A mixture of random models per request
- D. Human reviewers for all 10 million tickets
Answer: B. High volume + narrow task + cost/latency pressure → fine-tuned SLM, which also keeps data in-region. A is wrong: per-token SaaS pricing at 10M requests/month destroys the budget, and data leaves the region. C is nonsense. D doesn't scale.
How you did — and what's next
If you scored 8 or above, your agentic-architecture instincts are in good shape — but 10 questions don't cover the full exam. The real test goes deeper on ADK patterns, evaluation design, governance, and deployment trade-offs, all under time pressure. The way to close the gap is timed, full-length practice: take a complete exam, review every miss until you can explain why the right answer is right, map misses back to exam domains, and retake after 48 hours.
🎯 Want the full exam simulation?
These 10 questions are a warm-up. Our free Agentic Architect study guide covers every exam domain with flashcards and decision scenarios — and links the full-length timed practice tests on Udemy with the current discount applied.
Continue to the free study guide →
Comments
Post a Comment