CertifyAIxAce every AI certification.

NCP-AAI sample questions with answers

10 free practice questions for the NVIDIA-Certified Professional: Agentic AI exam. Try each one, then open the answer to see why the right option wins and every other option loses.

Question 1Agent Architecture and Design

An insurer's claims orchestrator runs as short-lived serverless functions. It delegates fraud reviews to a remote A2A agent, and a review can take several hours. The orchestrator cannot keep a connection open between invocations, and the platform team has banned tight polling loops because of cost. How should the orchestrator receive the review's result?

  1. A.

    Call the task retrieval operation every few seconds from a scheduled function until the task status reports completion

  2. B.

    Register a webhook through the task's push notification configuration so the remote agent calls back on significant updates

  3. C.

    Open a streaming request over server-sent events and keep it attached until the review task reaches a terminal state

  4. D.

    Send the review as a blocking request so that the remote agent returns only after the whole review has finished

Show answer

Answer: B

Push notifications fit very long tasks with clients that cannot hold a connection open.

  • A. Polling would technically work, but the stem bans tight polling loops, and polling adds latency and request cost for long tasks.
  • B. Correct: the remote agent calls the client's webhook on task updates, so no connection or polling is needed for an hours-long task.
  • C. Streaming needs the client to stay attached for the whole task, which a short-lived serverless function cannot do for hours.
  • D. A blocking call holds the request open until the task finishes, which fails for the same reason streaming does.
Question 2Agent Architecture and Design

A publisher's fact-checking agent answers questions that the model often knows but sometimes must verify with a search tool. The team wants to call the tool only when needed. When they sample several internal reasoning paths per question, the paths often agree on the same answer; when they disagree, the answer is usually wrong. What policy uses this signal?

  1. A.

    Always run the ReAct search loop first, and use sampled reasoning paths only to format the final answer

  2. B.

    Run the ReAct loop and the sampled reasoning paths in parallel on every question, and return whichever finishes first

  3. C.

    Always answer from a single greedy reasoning path, and use the search tool only when the user disputes the answer

  4. D.

    Answer from the sampled reasoning paths when a clear majority agrees, and fall back to a ReAct search loop when they do not

Show answer

Answer: D

Agreement among sampled reasoning paths signals whether internal knowledge suffices; when it is weak, back off to a tool-using ReAct loop.

  • A. Searching on every question ignores the stated goal of calling the tool only when needed, and it wastes the agreement signal.
  • B. Racing the methods pays for both on every question and selects by speed rather than by confidence.
  • C. One greedy path produces no agreement signal, and relying on user disputes pushes wrong answers to users first.
  • D. The ReAct authors propose backing off from self-consistency to ReAct when the majority answer is weak, signalling internal knowledge is insufficient.
Question 3Agent Architecture and Design

A freight broker is building a quoting agent. It must delegate customs-clearance questions to an agent that a partner firm operates and keeps opaque, and it must also run lookups against the broker's own tariff database. The team wants a standard protocol for each connection. Which protocol assignment fits these two connections?

  1. A.

    Use A2A to delegate work to the partner's clearance agent and MCP to expose the tariff database lookups as tools

  2. B.

    Use A2A for both connections, publishing the tariff database as an agent with its own Agent Card and skills

  3. C.

    Use MCP for both connections, registering the partner's clearance agent as one more tool with a fixed input schema

  4. D.

    Use MCP to reach the partner's clearance agent and A2A to reach the tariff database, since the database is internal

Show answer

Answer: A

A2A connects agents to other agents across a boundary; MCP connects an agent to its own tools and data.

  • A. Correct: agent-level delegation goes over A2A, and structured tool and data access goes over MCP, which is the documented split.
  • B. Wrapping a simple database lookup in a full A2A agent adds a reasoning layer and a task lifecycle that a structured query does not need.
  • C. MCP models discrete tools; forcing an opaque, multi-turn partner agent into a fixed tool schema loses task state and clarification exchanges.
  • D. This reverses the roles: whether the other side is internal or external does not decide the protocol, what it is (agent or tool) does.
Question 4Agent Architecture and Design

Harbourline Research runs an analyst agent over 40,000 internal strategy memos. Its vector RAG pipeline answers pointed questions well, but analysts increasingly ask things like "What are the main recurring risks across all our memos from the last three years?" Those questions name no company, product or person, and the answers returned cover only a handful of memos. Which retrieval approach addresses this class of question?

  1. A.

    Build an LLM-extracted knowledge graph, cluster it into communities, and answer from pre-generated community summaries

  2. B.

    Run a local graph search that starts at the entities named in the question and fans out to their neighbours

  3. C.

    Add a cross-encoder reranker after vector retrieval so that the most relevant memo chunks are promoted first

  4. D.

    Raise the vector search top-k substantially so that many more memo chunks reach the generator for each question

Show answer

Answer: A

Questions about themes across the whole corpus are what community-summary (global) search over an LLM-built knowledge graph is for.

  • A. Global, corpus-wide questions need a structure that already summarises the whole dataset; community summaries provide exactly that.
  • B. Local search needs a named entity to start from; these questions deliberately name none, so there is no seed to expand.
  • C. A reranker only reorders what retrieval already found; it cannot turn a few similar chunks into a corpus-wide synthesis.
  • D. More chunks still come from similarity to the query text, which points at nothing specific, and the context budget is soon exhausted.
Question 5Agent Architecture and Design

An insurer's claims agent uses a ReAct loop with a policy-search tool. On some claims, traces show the agent emitting the same thought and the same search call with identical arguments again and again until the request times out, although each search returns the same valid result. The team wants the loop to end predictably. Which change addresses this failure mode?

  1. A.

    Retry the policy-search tool automatically whenever it returns the same result twice, in case the index was refreshed between calls

  2. B.

    Increase the request timeout so the agent has more time to reach a final answer

  3. C.

    Cap the number of tool calls or iterations per request and stop with a fallback answer when the cap is reached

  4. D.

    Raise the sampling temperature so that each iteration produces a different search query and pushes the agent toward alternative reasoning paths

Show answer

Answer: C

A ReAct agent that repeats the same thought and action has failed to reason its way out of the loop, so the loop needs a hard iteration or tool-call bound.

  • A. The tool is returning a valid result; retrying it treats a reasoning failure as a tool failure and adds more calls.
  • B. A longer timeout only lets the same repetitive loop run longer and consume more tokens before failing.
  • C. A hard bound on iterations or tool calls is the documented guard against runaway loops, and it guarantees the request terminates.
  • D. Higher temperature adds randomness but no termination guarantee; the agent can still wander indefinitely, just less predictably.
Question 6Agent Architecture and Design

A utility's outage-response system has one supervisor agent directing 24 specialist agents covering grid telemetry, crews, customer messaging and regulators. As specialists were added, the supervisor's routing quality dropped and its context filled with status from every specialist. Leadership insists on keeping a clear chain of oversight. Which restructuring fits?

  1. A.

    Remove the supervisor and let all 24 specialists coordinate directly with each other as autonomous peers

  2. B.

    Introduce team-level supervisor agents beneath the top supervisor so each tier oversees a small group of specialists

  3. C.

    Replace the supervisor with a fixed sequential pipeline that runs the 24 specialists one after another

  4. D.

    Keep the single supervisor but give it a model with a larger context window so that all specialist status fits

Show answer

Answer: B

A hierarchy relieves an overloaded supervisor while preserving oversight.

  • A. Peer coordination removes the bottleneck but drops the clear oversight chain that leadership explicitly requires.
  • B. Correct: tiers of supervisors spread the coordination load and keep an explicit chain of oversight from top to specialists.
  • C. A fixed pipeline removes dynamic routing and serialises specialists that an outage needs working concurrently.
  • D. A bigger window only postpones the overload; accumulated context already degrades routing, and cost keeps growing with it.
Question 7Agent Architecture and Design

A finance-operations agent stores in long-term memory the note "Dana approved vendor payments up to the monthly limit". A reviewer finds that the payment tool executes transfers whenever such a memory entry is retrieved, without checking any other system. The memory store accepts writes derived from emails and chat messages. What exposure does this create?

  1. A.

    Retrieval latency may delay payments, because the tool waits for the memory search before each transfer

  2. B.

    Anyone who can influence what gets written to memory can obtain payment authority, because memory is treated as authorization

  3. C.

    The model could misspell the vendor name when writing the approval note, so payments go to the wrong vendor

  4. D.

    The memory store may run out of capacity as approvals accumulate, causing older approvals to be evicted

Show answer

Answer: B

Context can inform an action but must never authorize it; memory built from messages cannot stand in for an authorization check.

  • A. Latency is an operational concern, not an exposure; it does not let anyone obtain authority they should not have.
  • B. Memory derived from messages is model input, not trusted policy; using it as authorization lets crafted content trigger payments.
  • C. A spelling error is an accuracy defect in one entry, not a path for an outsider to authorize payments.
  • D. Capacity and eviction affect availability of approvals, not whether an attacker can create one.
Question 8Agent Architecture and Design

An engineering firm's chief-of-staff agent answers project-status questions from a memory of people, projects and decisions that it maintains from messages. When an answer is wrong, the team cannot tell whether the source messages, the memory update, the retrieval, or the model's final judgment was at fault. What should the memory design do?

  1. A.

    Keep source evidence separate from the derived knowledge the agent writes, with links from each derived entry to its sources

  2. B.

    Merge the evidence and the derived knowledge into a single document per project so that retrieval needs only one lookup

  3. C.

    Store only the raw source messages and let the model reinterpret them from scratch on every question

  4. D.

    Store only the derived knowledge and discard the source messages once they have been summarized

Show answer

Answer: A

Store derived knowledge alongside, not instead of, the evidence it came from, so errors can be traced to their stage.

  • A. Separating evidence from derived interpretation, with provenance, lets the team attribute an error to evidence, maintenance, retrieval or model.
  • B. Merging makes retrieval simpler but erases the boundary between what was said and what the agent concluded.
  • C. Raw-only storage avoids interpretation errors in memory but forces the agent to reconnect facts across time on every question.
  • D. Discarding sources removes the evidence needed to tell whether the summary or the source was wrong.
Question 9Agent Architecture and Design

A bank builds three agents for loan decisions: credit analysis, income verification and a decision agent. Each agent must see only the inputs it needs, and credit-bureau data must never reach the income agent. The decision agent needs both results. An architect proposes a shared blackboard that every agent reads and writes. Which communication design meets the requirement?

  1. A.

    Direct messages from each specialist to the decision agent, each payload scoped to the fields that the receiver needs

  2. B.

    The shared blackboard as proposed, since every agent then sees the full history, which aids auditability of the decision

  3. C.

    A group chat among all three agents so the decision agent can question both specialists in one accumulating thread

  4. D.

    The shared blackboard with a prompt telling the income agent to ignore any credit-bureau entries that it happens to read

Show answer

Answer: A

When agents must not see each other's inputs, scoped point-to-point messages beat a shared space every agent reads.

  • A. Correct: scoped point-to-point messages deliver both results to the decision agent without exposing bureau data to the income agent.
  • B. Auditability is a real benefit of a shared space, but every agent reading every entry breaks the stated confidentiality rule.
  • C. A group chat is a shared accumulating thread, so it has the same exposure problem as the proposed blackboard.
  • D. An instruction to ignore visible data is not an access control; the income agent still receives the credit-bureau entries.
Question 10Agent Architecture and Design

A retailer's order-support assistant handles one domain: order status, delivery changes and returns, using six well-described tools. There is no requirement for separate security boundaries, and the team does not need parallel work. An architect proposes splitting it into a supervisor plus six specialist agents for cleaner design. What should the team do?

  1. A.

    Keep a single agent with the six tools and iteration limits, since no tool overload, security boundary or parallelism justifies more agents

  2. B.

    Adopt the supervisor with six specialists, because one specialist per tool gives the most maintainable design

  3. C.

    Split it into two agents, one for reads and one for writes, because tool use always needs separate agents

  4. D.

    Adopt a group chat of six specialists, because consensus between agents raises the accuracy of every answer

Show answer

Answer: A

Use the lowest level of complexity that meets the need; a single tool-using agent often suffices.

  • A. Correct: one domain, few tools, no security boundaries and no parallelism give no reason to pay multi-agent overhead.
  • B. Specialization helps maintainability in some systems, but one agent per tool here only adds hops, latency and failure modes.
  • C. No rule requires separate agents for reads and writes; that split adds a handoff without a stated need.
  • D. A consensus discussion adds latency and cost to routine lookups and needs a chat manager to control turn-taking.

Keep going with 623 more NCP-AAI questions

Free papers every day, in the real exam formats, with progress by exam domain. Unlock every paper and timed mock exam when you are ready.

Practise freeSee all NCP-AAI papers
NCP-AAI sample questions with answers (10 free) · CertifyAIx