NVIDIA NCP-AAI: NVIDIA-Certified Professional: Agentic AI

Last checked October 22, 2026 against the official NVIDIA exam page.

1. Question 1Agent Architecture and Design

   An LLM agent must decide which tool to call, observe the result, and then decide the next step until a task is complete. Which pattern describes this loop?

   1. AStatic retrieval with a single response over a frozen corpus
   2. BReAct, interleaving reasoning steps with tool actions and observations
   3. CBatch normalization with a fixed schedule, applied per layer
   4. DPure next-token completion with no tool actions in the loop
   Show answer and explanationHide answer

   Correct answer: B

   ReAct alternates reasoning with actions and observations, which is the standard loop for tool-using agents. Batch normalization is a training technique, pure completion has no action loop, and single retrieval does not iterate.

2. Question 2Tool Use and Integration

   An agent must call an internal ticketing API that requires a token. Where should the credential be stored?

   1. AInside the prompt so the model can see it, where it is echoed in traces
   2. BIn the model's training data, ingested during pretraining and fine-tuning
   3. CIn a secrets manager, with the agent retrieving it at run time under a scoped identity
   4. DIn the public repository's README, committed with the agent source code
   Show answer and explanationHide answer

   Correct answer: C

   A secrets manager with scoped identities keeps credentials out of prompts, code and training data. Prompts, public READMEs and training data expose the token.

3. Question 3Safety and Guardrails

   A user attempts to make an agent reveal its system instructions by asking it to ignore prior rules. Which defense is most effective?

   1. AGiving the agent admin rights, so it can override the policies set by the operator and bypass the refusal policy
   2. BSeparating system instructions from user input and using guardrails that refuse disclosure requests
   3. CPrinting the system instructions on the homepage, where visitors can read the text
   4. DAllowing the user to edit the system instructions, so the prompt can be rewritten
   Show answer and explanationHide answer

   Correct answer: B

   Separation of instructions and guardrails that refuse disclosure protect the system prompt. Publishing instructions, granting admin rights and letting users edit instructions all weaken control.

4. Question 4Evaluation and Observability

   Which metric best measures whether an agent completed a user's task successfully across a test set?

   1. ANumber of tokens generated per second, averaged across the decoding steps
   2. BTask success rate evaluated against defined success criteria
   3. CCount of tools available in the registry, including deprecated tools
   4. DNumber of log lines written, emitted at debug level by each module
   Show answer and explanationHide answer

   Correct answer: B

   Task success rate against defined criteria measures outcomes that matter to users. Token speed, tool counts and log volume measure activity or performance rather than success.

5. Question 5NVIDIA Agent Tooling

   Which NVIDIA toolkit helps build, profile, and evaluate agent workflows, including connections to tools and models?

   1. ANVIDIA Omniverse Nucleus
   2. BNVIDIA DeepStream SDK
   3. CNVIDIA NeMo Agent Toolkit
   4. DNVIDIA Clara Holoscan
   Show answer and explanationHide answer

   Correct answer: C

   The NeMo Agent Toolkit supports building and evaluating agent workflows with tools and models. Omniverse is for simulation collaboration, DeepStream for video analytics and Holoscan for medical and sensor applications.