I85PREP
⌘ K
NVIDIA NCP-AAI study materials
NVIDIA NCP-AAI study resources

NVIDIA NCP-AAI practice questions

One question at a time. Build your NVIDIA NCP-AAI understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Agentic AI solution design and multi-agent interaction

In an agentic AI system built with NVIDIA NIM microservices, which component is primarily responsible for deciding the next action based on the conversation history and available tools?

  1. The tokenizer inside the inference server
  2. The load balancer fronting the microservices
  3. The LLM acting as the reasoning core of the agent Correct answer.
  4. The vector database storing embedded documents

Explanation

In an agentic architecture the large language model is the reasoning core: it reads the accumulated context, reasons about the goal, and selects the next action or tool call. Retrieval stores and infrastructure components only supply context or transport requests; they do not perform planning.

Question 2Agentic AI solution design and multi-agent interaction

A developer wants an agent to call an external weather API reliably and return structured arguments. Which mechanism should the agent use to request that call?

  1. A structured tool or function call with typed arguments Correct answer.
  2. A free-text paragraph describing the desired forecast
  3. A raw SQL query embedded in the system prompt
  4. A streaming token continuation of the user message

Explanation

Tool or function calling lets the model emit a structured request naming the tool and supplying typed arguments that the runtime can validate and execute. This is more reliable than parsing free text and gives the orchestrator a deterministic contract for invoking external APIs.

Question 3Agentic AI solution design and multi-agent interaction

Which pattern best describes a supervisor agent coordinating several specialized worker agents in an NVIDIA agentic workflow?

  1. A flat peer network where every agent broadcasts to every other agent
  2. A stateless request pipeline with no shared conversation state
  3. A single monolithic prompt that contains every worker's instructions
  4. A hierarchical orchestration where the supervisor routes tasks and aggregates results Correct answer.

Explanation

Supervisor-worker designs are hierarchical: the supervisor decomposes the goal, routes subtasks to specialized workers, and merges their outputs into a final answer. This keeps each worker's prompt focused and makes routing decisions explicit and auditable.

Question 4Agentic AI solution design and multi-agent interaction

An agent must answer questions over a large internal knowledge base that changes daily. Which approach keeps responses grounded without retraining the model?

  1. Shorten the system prompt to leave room for memorized knowledge
  2. Retrieve relevant chunks at query time and inject them into the prompt Correct answer.
  3. Increase the model's temperature so it recalls fresher facts
  4. Fine-tune the model every night on the new documents

Explanation

Retrieval-augmented generation fetches the most relevant chunks for each query and places them in the model context, so answers reflect current content without weight updates. This is cheaper and faster to refresh than nightly fine-tuning and keeps claims traceable to sources.

Question 5Agentic AI solution design and multi-agent interaction

What is the primary purpose of maintaining explicit conversation state in a multi-turn agent?

  1. To reduce the model's parameter count between turns
  2. To eliminate the need for any tool definitions
  3. To guarantee deterministic sampling across all requests
  4. To let the agent reference earlier results and avoid repeating completed steps Correct answer.

Explanation

Explicit state records prior messages, tool results, and decisions so the agent can build on earlier work instead of restarting. Without it, multi-step tasks lose continuity and the agent may repeat calls or contradict its own findings.

Question 6Agentic AI solution design and multi-agent interaction

In an agentic RAG pipeline, which component decides whether a retrieved passage is relevant enough to answer the user's question?

  1. The GPU driver managing inference memory
  2. The embedding model that produced the passage vector
  3. The document ingestion scheduler
  4. A relevance check or reranking step before generation Correct answer.

Explanation

A relevance check or reranker scores retrieved passages against the query and filters weak matches before they reach the generator. This reduces hallucination grounded in irrelevant text and is a distinct step from embedding and ingestion.

Question 7Agentic AI solution design and multi-agent interaction

Which NVIDIA offering provides a catalog of prebuilt, optimized inference microservices that agents can call for models such as LLMs and embedding models?

  1. NVIDIA CUDA Toolkit
  2. NVIDIA DOCA
  3. NVIDIA TensorRT-LLM compiler only
  4. NVIDIA NIM Correct answer.

Explanation

NVIDIA NIM is the catalog of containerized inference microservices exposing standardized APIs for LLMs, embedding models, and rerankers. Agents and orchestrators call these endpoints instead of managing model serving internals directly.

Question 8Agentic AI solution design and multi-agent interaction

When an agent's tool call fails with a transient timeout, which behavior is most appropriate for a resilient agent loop?

  1. Loop indefinitely until the tool eventually responds
  2. Silently invent a plausible result to keep the conversation moving
  3. Retry with a bounded backoff and then fall back or report the failure Correct answer.
  4. Immediately return the raw error to the user without retrying

Explanation

Resilient agents retry transient failures a bounded number of times with backoff, then escalate or fall back when the limit is reached. This balances recovery against runaway loops and avoids fabricating data that the user would trust.

Question 9Agentic AI solution design and multi-agent interaction

A team wants two agents to collaborate on a coding task: one writes code and one reviews it. Which interaction design most directly supports this?

  1. The reviewer agent is given no access to the writer's output
  2. The writer passes its output to the reviewer, whose feedback is returned to the writer Correct answer.
  3. Both agents share a single system prompt and run once concurrently
  4. The writer agent calls the reviewer only after the user has approved the code

Explanation

A generate-then-review loop requires the writer's artifact to flow to the reviewer and the reviewer's critique to return to the writer for revision. This explicit handoff and feedback path is what enables iterative improvement between specialized agents.

Question 10Agentic AI solution design and multi-agent interaction

Which situation most strongly favors splitting a monolithic agent into multiple specialized agents?

  1. The team wants to minimize the number of API calls per request
  2. The task is a single classification of one short sentence
  3. The task has distinct phases requiring different tools, prompts, and validation criteria Correct answer.
  4. The latency budget is extremely tight for one model call

Explanation

Specialization pays off when phases differ in tools, instructions, or acceptance criteria, because each agent can carry a focused prompt and narrower toolset. Simple single-step tasks or tight latency budgets usually favor one agent, since multi-agent handoffs add calls and coordination cost.

Question 11Agentic AI solution design and multi-agent interaction

An agent is given a tool that can execute shell commands. Which safeguard is most important before allowing it to run autonomously?

  1. Increasing the model temperature to encourage caution
  2. Logging only the final answer returned to the user
  3. Restricting the tool's permissions and sandboxing its execution environment Correct answer.
  4. Adding more examples of shell commands to the system prompt

Explanation

Tools with real side effects must be constrained by least-privilege permissions and sandboxing so a mistaken or manipulated call cannot damage systems. Prompt examples and temperature do not provide enforcement, and logging only final answers removes the audit trail needed to detect misuse.

Question 12Agentic AI solution design and multi-agent interaction

In a multi-agent system, agent A's output is inserted directly into agent B's prompt. Which risk does this introduce?

  1. A drop in the embedding dimensionality of agent B
  2. Prompt injection or instruction leakage propagating between agents Correct answer.
  3. Automatic recompilation of the underlying model weights
  4. Loss of GPU parallelism across the inference cluster

Explanation

Inter-agent message passing carries untrusted text into another agent's context, so malicious or garbled instructions can propagate and hijack downstream behavior. Designs should treat agent outputs as untrusted data, separate them from instructions, and validate them before reuse.

Question 13Agentic AI solution design and multi-agent interaction

A retrieval agent keeps returning irrelevant chunks because it always fetches the top three results regardless of score. Which change best improves answer quality?

  1. Apply a similarity threshold and rerank before passing chunks to the generator Correct answer.
  2. Increase the generator's maximum output tokens
  3. Always retrieve the top fifty chunks to maximize coverage
  4. Remove the embedding step and pass the whole corpus to the model

Explanation

Filtering by a relevance threshold and reranking keeps only passages that actually support the answer, reducing noise in the generator's context. Retrieving more chunks or the full corpus adds distraction, and output length does not fix bad grounding.

Question 14Agentic AI solution design and multi-agent interaction

Which metric best captures whether a multi-agent pipeline's final answers are supported by the retrieved evidence?

  1. Total GPU memory reserved by the inference server
  2. Number of agents in the topology
  3. Groundedness or faithfulness of the final response to cited sources Correct answer.
  4. Average tokens per agent call

Explanation

Groundedness or faithfulness measures whether claims in the final answer are supported by the retrieved sources, which is the core quality goal of a RAG-based multi-agent pipeline. Token counts, agent counts, and memory use describe cost and architecture, not answer correctness.

Question 15Agentic AI solution design and multi-agent interaction

An agent must choose between three tools with overlapping capabilities. Which design most reduces ambiguous tool selection?

  1. Remove all tool descriptions so the model must infer usage
  2. Merge the three tools into one with no documented parameters
  3. Give each tool a clear, narrowly scoped description stating when to use it Correct answer.
  4. Add all three tools to every prompt with identical descriptions

Explanation

Tool selection is driven largely by the names, descriptions, and parameter schemas the model sees, so narrow and distinct descriptions help the model route correctly. Removing or duplicating descriptions increases ambiguity and raises the chance of wrong or redundant calls.

Question 16Agentic AI solution design and multi-agent interaction

What is the main benefit of streaming intermediate agent steps, such as tool calls and partial results, to the client?

  1. It guarantees the final answer will be factually correct
  2. It improves perceived responsiveness and gives visibility into the agent's reasoning path Correct answer.
  3. It removes the need for a conversation state store
  4. It reduces the model's parameter count during generation

Explanation

Streaming intermediate events lets users see progress during long multi-step tasks and makes the agent's behavior auditable. It improves transparency and perceived latency but does not change model size, guarantee correctness, or replace persistent state.

Question 17Agentic AI solution design and multi-agent interaction

A planner agent produces a multi-step plan, but the executor agent must handle a step whose tool is unavailable. Which interaction pattern is most robust?

  1. The executor reports the obstacle back to the planner for replanning Correct answer.
  2. The executor silently skips the step and reports success
  3. The executor substitutes an unrelated available tool without informing the planner
  4. The executor restarts the entire plan from the first step

Explanation

Returning the obstacle to the planner lets the system revise the plan with full knowledge of the goal and constraints. Silent skips hide failures, full restarts waste completed work, and unannounced substitutions can produce results that do not satisfy the original intent.

Question 18Agentic AI solution design and multi-agent interaction

Which practice best keeps a long-running agent within a model's context window while preserving important earlier findings?

  1. Summarize or checkpoint key results and retrieve them when needed Correct answer.
  2. Increase the temperature to compress the context
  3. Send the full unbounded history on every request
  4. Discard the oldest messages unconditionally at every turn

Explanation

Summarization and checkpointing condense important findings into compact state that can be re-injected or retrieved later, keeping the working context within limits without losing critical facts. Blind truncation loses information, and unbounded history eventually exceeds the window.

Question 19Agentic AI solution design and multi-agent interaction

An agent is asked to summarize a document while also calling a retrieval tool. Which orchestration approach lets it decide the order of these actions dynamically?

  1. A single non-tool model call with no further steps
  2. A fixed pipeline that always retrieves before reading the document
  3. An agent loop where the model selects the next action from available tools until the goal is met Correct answer.
  4. A hardcoded sequence that never changes across requests

Explanation

An agent loop lets the model choose among tools and actions based on current context, so the order can adapt to the document and the question. Fixed pipelines and single calls cannot reorder steps in response to intermediate results.

Question 20Agentic AI solution design and multi-agent interaction

Which element is essential for an agent to recover gracefully after producing an invalid tool-call payload?

  1. A second model with a higher temperature
  2. A schema validator that returns a descriptive error the agent can act on Correct answer.
  3. Removing the tool from the registry permanently
  4. A larger maximum token limit for the final answer

Explanation

Validating tool arguments against a schema and returning a specific error message lets the agent correct its payload on the next attempt. Token limits and temperature do not diagnose malformed arguments, and removing the tool prevents completing the task.

Question 21Agentic AI solution design and multi-agent interaction

A team observes that adding more worker agents to a supervisor system increased latency without improving answer quality. Which is the most likely explanation?

  1. The supervisor's temperature was set to zero
  2. The GPU cluster lost its network connection
  3. The embedding model was silently upgraded
  4. Coordination overhead and redundant calls grew faster than any quality benefit Correct answer.

Explanation

Each additional worker adds routing decisions, handoffs, and often redundant model calls, so latency can rise without quality gains when workers do not contribute distinct value. This is a common signal to consolidate agents or tighten routing criteria.

Question 22Agentic AI solution design and multi-agent interaction

An agent must answer questions about a policy that changes weekly and must cite the exact clause. Which combination best satisfies both freshness and traceability?

  1. Raise temperature so the model paraphrases clauses more creatively
  2. Fine-tune weekly and ask the model to recall clause numbers from memory
  3. Retrieve current policy chunks and require the agent to cite the retrieved chunk identifiers Correct answer.
  4. Cache answers from the first week and reuse them indefinitely

Explanation

Retrieving the current policy text ensures freshness, and requiring citations to retrieved chunk identifiers gives traceability back to the source. Fine-tuning from memory, stale caching, and creative paraphrasing all undermine accuracy and verifiability.

Question 23Agentic AI solution design and multi-agent interaction

A multi-agent system allows a research agent to write files that a deployment agent later executes. Which control most directly limits the blast radius of a compromised research agent?

  1. Enforce least-privilege credentials and require human approval for execution steps Correct answer.
  2. Log the research agent's outputs only after deployment completes
  3. Increase the research agent's context window
  4. Give the research agent administrator rights so it can fix its own errors

Explanation

Least-privilege credentials and human approval gates on high-impact actions such as execution constrain what a compromised or mistaken agent can do. Broad privileges enlarge the blast radius, and post-deployment logging is too late to prevent damage.

Question 24Agentic AI solution design and multi-agent interaction

A supervisor routes tasks to workers, but workers sometimes return confident yet incorrect results. Which design change most improves overall reliability?

  1. Increase the number of workers so at least one is likely correct
  2. Add a verification step that checks worker outputs against evidence before aggregation Correct answer.
  3. Remove the supervisor so workers respond directly to the user
  4. Have the supervisor accept the first worker response to minimize latency

Explanation

Verification against retrieved evidence or explicit criteria catches confident errors before they are aggregated into the final answer. Accepting the first response, adding unverified workers, or removing the supervisor all increase the chance that incorrect outputs reach the user.

Question 25Agentic AI solution design and multi-agent interaction

An agentic system must handle a task where the correct sequence of tool calls is unknown in advance and depends on intermediate results. Which failure mode is most important to design against?

  1. The embedding index being stored on local disk
  2. The agent looping without progress or terminating prematurely on a wrong assumption Correct answer.
  3. The model's vocabulary size being too small
  4. The inference server using a single GPU

Explanation

When the path is discovered at runtime, agents can cycle through the same actions without progress or stop early on an incorrect assumption, so designs need step limits, progress checks, and state tracking. Vocabulary size, index location, and GPU count are deployment details that do not address this control-flow risk.

Question 26Agentic AI solution design and multi-agent interaction

In an agentic AI system, which component is responsible for deciding the next action at each step of a reasoning loop?

  1. The observability exporter
  2. The large language model acting as the reasoning core Correct answer.
  3. The API gateway fronting the tools
  4. The vector database storing embeddings

Explanation

In an agentic loop, the LLM itself performs the reasoning that selects the next action, such as calling a tool, refining a plan, or producing a final answer. The surrounding infrastructure only supplies context and executes the model's chosen actions.

Question 27Agentic AI solution design and multi-agent interaction

Which pattern describes a single agent that repeatedly calls tools and observes results until it produces a final answer?

  1. One-shot prompt completion
  2. ReAct-style reasoning and acting loop Correct answer.
  3. Static batch inference
  4. Fine-tuning with supervised labels

Explanation

The ReAct pattern interleaves reasoning traces with tool actions and observations, letting the agent iterate toward a goal. Static inference, one-shot completion, and fine-tuning do not involve iterative tool use.

Question 28Agentic AI solution design and multi-agent interaction

What is the primary purpose of defining an explicit JSON schema for a tool that an agent can call?

  1. To guarantee the tool's backend will never return an error
  2. To encrypt the tool's responses before they reach the agent
  3. To constrain the arguments the model must produce so calls are machine-parseable and valid Correct answer.
  4. To reduce the model's token cost by shortening the system prompt

Explanation

A tool schema declares required fields, types, and constraints so the model emits structured arguments that the runtime can validate and execute. It improves reliability of tool invocation but does not eliminate backend errors or encrypt traffic.

Question 29Agentic AI solution design and multi-agent interaction

A team wants an agent to answer questions over internal documents using retrieval, and it must also be able to look up live order status through an API. Which design best fits these requirements?

  1. A single static prompt containing all documents and order records
  2. An agent with a retrieval tool and an order-status API tool, choosing between them per query Correct answer.
  3. A retrieval-only agent that ignores order-status questions
  4. A fine-tuned model trained only on order-status data

Explanation

Giving the agent distinct tools for retrieval and live API access lets it select the right capability for each question. Static prompts cannot hold live data, and single-purpose designs fail one of the two requirements.

Question 30Agentic AI solution design and multi-agent interaction

Which statement best describes the role of a supervisor agent in a multi-agent system?

  1. It executes every subtask itself without delegation
  2. It stores embeddings for all worker agents
  3. It replaces the need for any tool definitions
  4. It decomposes the goal, assigns subtasks to workers, and consolidates their results Correct answer.

Explanation

A supervisor coordinates a multi-agent system by planning, routing subtasks to specialized workers, and merging or validating their outputs. It does not replace tool definitions or act as an embedding store.

Question 31Agentic AI solution design and multi-agent interaction

Why is it important for an agent's tool-call output to be validated before execution?

  1. Validation guarantees the tool's business logic is correct
  2. Malformed or unexpected arguments can cause unsafe or incorrect actions Correct answer.
  3. Validation removes the need for any access control
  4. Validation makes the model produce shorter answers

Explanation

Models can emit invalid JSON, missing fields, or out-of-range values, so a validation layer prevents unsafe or erroneous execution. Validation does not replace access control or verify business logic.

Question 32Agentic AI solution design and multi-agent interaction

Which NVIDIA component is designed to provide GPU-accelerated building blocks for retrieval pipelines used by agents?

  1. NVIDIA DRIVE
  2. NVIDIA Clara
  3. NVIDIA Omniverse
  4. NVIDIA NeMo Retriever microservices Correct answer.

Explanation

NeMo Retriever provides GPU-accelerated microservices for embedding, indexing, and reranking that support retrieval-augmented agent workflows. Omniverse targets simulation, DRIVE targets automotive, and Clara targets healthcare.

Question 33Agentic AI solution design and multi-agent interaction

An agent must summarize a long report and also fetch related statistics from an external database. Which capability is essential for handling both in one session?

  1. The ability to run only a single fixed prompt
  2. The ability to disable all external tools
  3. The ability to interleave reasoning with tool calls and incorporate observations Correct answer.
  4. The ability to discard intermediate results after each step

Explanation

Interleaving reasoning with tool calls lets the agent fetch statistics and then fold those observations into the summary. Fixed prompts and discarding results prevent combining external data with the document.

Question 34Agentic AI solution design and multi-agent interaction

What distinguishes a multi-agent architecture from a single-agent architecture with multiple tools?

  1. The system streams tokens to the client
  2. Work is divided among separate agent instances with their own roles and context Correct answer.
  3. The system uses a larger language model
  4. The system uses more than one tool definition

Explanation

Multi-agent designs decompose work across distinct agents that each maintain their own role, instructions, and context, coordinating through messages or a supervisor. Merely having multiple tools, a larger model, or streaming does not make an architecture multi-agent.

Question 35Agentic AI solution design and multi-agent interaction

A planner agent produces a task graph, and several worker agents execute nodes concurrently. Which mechanism best prevents two workers from overwriting each other's shared state?

  1. Letting each worker write freely and resolve conflicts later
  2. Removing shared state and giving workers no memory
  3. A coordination layer that serializes or locks writes to shared state Correct answer.
  4. Increasing the number of workers to outpace conflicts

Explanation

Concurrent writers to shared state need coordination such as locking, versioning, or a single writer to avoid lost updates. Free-for-all writes and adding more workers increase the chance of conflicting overwrites.

Question 36Agentic AI solution design and multi-agent interaction

Which approach best enables an agent to use a newly added tool without retraining the underlying model?

  1. Fine-tuning the model on examples of the new tool
  2. Retraining the model from scratch on all tools
  3. Hard-coding the tool call into the model weights
  4. Describing the tool in the prompt with a schema and letting the model select it at inference time Correct answer.

Explanation

Tool use through in-context descriptions and schemas lets a model invoke new tools at inference time without weight updates. Fine-tuning or retraining is slower and unnecessary for adding a single tool.

Question 37Agentic AI solution design and multi-agent interaction

An agent's plan requires a subtask that no available worker agent is equipped to handle. Which response is most appropriate for the supervisor?

  1. Abort the entire workflow without reporting the gap
  2. Silently skip the subtask and report success
  3. Re-plan to decompose the subtask further or escalate to a fallback path Correct answer.
  4. Assign the subtask to a random worker regardless of capability

Explanation

When no worker matches a subtask, the supervisor should re-plan, decompose further, or escalate so the gap is addressed transparently. Skipping, random assignment, or silent aborts produce incorrect or opaque outcomes.

Question 38Agentic AI solution design and multi-agent interaction

Which design most directly reduces the risk of an agent entering an infinite loop of repeated identical tool calls?

  1. Removing the tool result from the context
  2. Giving the agent more tools to choose from
  3. Tracking call history and enforcing a maximum iteration or repetition limit Correct answer.
  4. Increasing the model's temperature on every call

Explanation

Detecting repeated calls and enforcing iteration limits stops runaway loops. Raising temperature, adding tools, or hiding results does not address the repetition and may worsen behavior.

Question 39Agentic AI solution design and multi-agent interaction

In a multi-agent system, what is the main purpose of passing structured messages between agents rather than free-form text?

  1. Structured messages make intent and payloads easier to validate and route reliably Correct answer.
  2. Structured messages eliminate the need for any model reasoning
  3. Structured messages guarantee that agents never disagree
  4. Structured messages reduce the number of agents required

Explanation

Structured messages with defined fields let the system validate content and route it to the right agent, improving reliability. They do not remove reasoning needs, guarantee agreement, or reduce agent count.

Question 40Agentic AI solution design and multi-agent interaction

An agent must call a tool that occasionally returns partial data. Which design best lets the agent handle partial results correctly?

  1. Assume every response is complete and proceed
  2. Disable the tool after the first partial response
  3. Instruct the agent to check the response status and retry or supplement when data is incomplete Correct answer.
  4. Ignore the status field and parse whatever is present

Explanation

Explicit status checks let the agent detect incompleteness and retry or supplement before proceeding. Assuming completeness or ignoring status leads to incorrect downstream results.

Question 41Agentic AI solution design and multi-agent interaction

Which consideration most affects how many specialized agents a multi-agent design should include?

  1. The length of the model's system prompt only
  2. The color scheme of the user interface
  3. The number of GPUs available for inference
  4. The trade-off between task specialization gains and added coordination overhead Correct answer.

Explanation

Each added agent brings specialization benefits but also coordination, latency, and failure-mode costs, so the right count balances the two. Hardware count, UI styling, and prompt length alone are not the deciding factors.

Question 42Agentic AI solution design and multi-agent interaction

What is the benefit of giving each worker agent a narrow, well-scoped system prompt in a multi-agent system?

  1. It makes the supervisor unnecessary
  2. It allows the agent to ignore its tools
  3. It focuses the agent's behavior and reduces unintended tool use Correct answer.
  4. It guarantees deterministic outputs

Explanation

Narrow prompts keep each worker focused on its role, reducing drift and unintended tool calls. They do not remove the supervisor, let agents ignore tools, or guarantee determinism.

Question 43Agentic AI solution design and multi-agent interaction

Which pattern best supports an agent that must verify its own answer against a source before responding to the user?

  1. A critic or verification step that checks the draft answer against retrieved evidence Correct answer.
  2. Removing retrieval to speed up the response
  3. A higher temperature setting for creativity
  4. A single generation pass with no review

Explanation

A verification step, often implemented as a critic agent or self-check, compares the draft against evidence before release. Single-pass generation, higher temperature, or removing retrieval weakens grounding.

Question 44Agentic AI solution design and multi-agent interaction

A team observes that a multi-agent pipeline produces inconsistent answers across runs for identical inputs. Which factor most likely contributes?

  1. The use of structured messages
  2. The presence of a validation layer
  3. The use of a fixed tool schema
  4. Non-deterministic sampling and variable message ordering among agents Correct answer.

Explanation

Sampling randomness and nondeterministic ordering of concurrent agent messages can cause run-to-run variation. Fixed schemas, validation, and structured messages generally improve consistency rather than reduce it.

Question 45Agentic AI solution design and multi-agent interaction

Which design best allows a supervisor to recover when a worker returns an error status instead of a result?

  1. The supervisor ignores status codes entirely
  2. The supervisor treats the error as a valid result and continues
  3. The supervisor terminates the whole system on any worker error
  4. The supervisor inspects the status, retries or reassigns the subtask, and records the failure Correct answer.

Explanation

Status-aware handling lets the supervisor retry, reassign, or escalate while logging the failure for diagnosis. Treating errors as results or terminating on any error harms robustness.

Question 46Agentic AI solution design and multi-agent interaction

An agent is designed to draft a contract clause by clause, consulting a legal knowledge base for each clause. Which orchestration approach best fits this task?

  1. A single prompt that asks for the entire contract at once with no retrieval
  2. A tool that executes the contract automatically
  3. A plan-and-execute loop that iterates over clauses, retrieving evidence per clause Correct answer.
  4. A retrieval step performed only once before drafting

Explanation

Iterating clause by clause with per-clause retrieval grounds each part in relevant evidence, which suits structured multi-step drafting. A single ungrounded pass or one upfront retrieval risks unsupported clauses.

Question 47Agentic AI solution design and multi-agent interaction

A multi-agent system routes each user request to one of several specialist agents. Which failure mode is most important to detect when specialists have overlapping domains?

  1. Excessive token streaming to the client
  2. Presence of a logging layer
  3. Use of a schema for tool arguments
  4. Misrouting a request to a specialist whose answer is plausible but out of scope Correct answer.

Explanation

Overlapping specialties can produce confident but out-of-scope answers when routing is wrong, so routing accuracy and scope checks matter most. Streaming, schemas, and logging are not routing failure modes.

Question 48Agentic AI solution design and multi-agent interaction

Which architecture best supports an agent that must maintain a coherent long conversation while delegating research to sub-agents?

  1. Sub-agents that each maintain independent copies of the full conversation
  2. A primary agent that owns the conversation state and spawns stateless research sub-agents per query Correct answer.
  3. A single agent that performs all research inline without delegation
  4. A stateless primary agent that discards history after each turn

Explanation

Keeping conversation state in the primary agent while delegating focused research to sub-agents preserves coherence and avoids duplicating full history across workers. Inline research or stateless designs sacrifice scalability or continuity.

Question 49Agentic AI solution design and multi-agent interaction

A team wants to add a new specialist agent to an existing supervisor-worker system without breaking current routing. Which change is safest?

  1. Remove capability descriptions so routing is implicit
  2. Route all requests to the new agent by default
  3. Register the new agent with a distinct capability description and update routing rules with tests Correct answer.
  4. Replace the supervisor with a single monolithic agent

Explanation

Registering a distinct capability and updating routing with tests lets the supervisor select the new agent without disturbing existing paths. Removing descriptions or defaulting all traffic to the new agent destabilizes routing.

Question 50Agentic AI solution design and multi-agent interaction

Which evaluation approach best measures whether a multi-agent pipeline's decomposition improves outcomes over a single-agent baseline?

  1. Compare the size of the system prompts
  2. Compare task success and cost on a held-out benchmark across both architectures with identical tools Correct answer.
  3. Count the number of agents in each design
  4. Measure only the latency of the multi-agent pipeline

Explanation

A controlled comparison on the same benchmark with identical tools isolates the effect of decomposition on success and cost. Latency alone, agent counts, or prompt sizes do not measure outcome quality.