NCP-AAI mock exam

Question 1, Agent Architecture and DesignWhich situation best justifies an agent over a fixed workflow?

• AThe steps never change, but the team expects that they may one day
• BThe task is a single classification with a large number of labels
• CLatency must be as low as possible on every request
• DThe path to the answer depends on what is found along the way

Question 2, Agent DevelopmentAn agent's tool calls often fail on bad arguments. What improves this most?

• AA bigger GPU so the model can reason over more tool options
• BPrecise tool schemas and descriptions, with validation
• CMore agents in the system so that each can check the others
• DA higher temperature so the model tries other arguments

Question 3, Cognition, Planning, and MemoryAn agent must recall user preferences across separate sessions. Which memory type fits?

• AThe tokenizer's session cache
• BThe context window of each session
• CThe KV cache on the serving GPU
• DLong-term memory in an external store

Question 4, Cognition, Planning, and MemoryA plan fails at step three because a tool result was unexpected. What should a good agent do?

• ARepeat step three until the tool returns the expected result
• BRevise the plan based on the new information
• CFinish the remaining steps anyway and flag the problem at the end
• DStop and delete all progress so the next run starts clean

Question 5, Evaluation and TuningYou want to know whether a new agent version is better. What should you measure?

• AThe number of tokens used per task, compared across the two versions
• BDeveloper opinion after a demo of each agent version
• CThe size of the system prompt and the number of tools
• DTask success on a fixed scenario set, plus tool accuracy and cost

Question 6, Deployment and ScalingA multi-agent service must handle growing load. Where should you look first when scaling?

• AAdding more context to each prompt to reduce repeated tool calls
• BRemoving logging and tracing to lower the overhead per request
• CStateless components, external state, and queueing for long tasks
• DStoring session state in each process's memory for faster access

Question 7, NVIDIA Platform ImplementationYou want to compose, profile and evaluate agents built with different frameworks. What is the NVIDIA tool?

• ANVIDIA Omniverse Kit SDK
• BNVIDIA NeMo Agent Toolkit
• CNVIDIA RAPIDS Accelerator
• DNVIDIA DCGM Exporter

Question 8, Human-AI Interaction and OversightAn agent can delete customer accounts. What is the right oversight?

• AAllow the action only during business hours when staff are around
• BRequire human approval before the action runs
• CAutomatic execution, with a log that is reviewed each week
• DA clear warning in the system prompt about deleting accounts