Agentic Architecture & Orchestration Quiz - Certifications Claude Architect Foundations
Free practice quiz: Agentic Architecture & Orchestration (Certifications Claude Architect Foundations). 73 multiple-choice questions with answers and explanations. Play online, no sign-up needed.
Play all 73 questions freeSample questions
What is the best fix to make the loop continue when the model still needs tools?
- Add a prompt that forbids writing I am going to before all required tool calls are made
- Increase the iteration cap so the loop cannot stop before the refund workflow completes
- Use stop_reason to continue on tool_use and stop on end_turn or terminal reasons
- Put tool_result blocks before the assistant message instead of in a new user message
Show answer
Answer: C. Use stop_reason to continue on tool_use and stop on end_turn or terminal reasons
The agentic loop is controlled by stop_reason. Checking natural language text for completion is an anti-pattern; the loop should continue while stop_reason is tool_use and exit on terminal reasons such as end_turn.
What is the best coordinator change to improve first-contact resolution while preserving the hub-and-spoke architecture?
- Allow the refund-eligibility subagent to escalate directly to a human when REFUND_DENIED occurs so the other subagents can still continue running independently.
- Increase the coordinator's context window size so that it can log all three subagent transcripts before aborting the entire current support ticket.
- Handle REFUND_DENIED as a normal subagent result, combine successful profile and order data, and escalate only the refund decision requiring human approval.
- Remove the refund-eligibility subagent and instead have the profile-lookup subagent call process_refund directly so fewer handoffs can cause a ticket to abort.
Show answer
Answer: C. Handle REFUND_DENIED as a normal subagent result, combine successful profile and order data, and escalate only the refund decision requiring human approval.
A coordinator should treat a subagent error as one result to aggregate, not as a signal to discard successful results from other subagents. Routing error handling through the coordinator lets it escalate only the refund decision that requires human approval while still using the profile and order results. This preserves controlled information flow and consistent error handling.
What is the most likely root cause?
- The synthesis subagent's AgentDefinition.tools does not include Task, so it cannot spawn the fact-checker.
- The fact-checker's description is too specific, so the synthesis subagent decides not to delegate to it.
- The coordinator's system prompt is too restrictive, which prevents the synthesis subagent from selecting tools.
- The fact-checker subagent is not listed in the synthesis subagent's own agents field, so the synthesis subagent cannot see it.
Show answer
Answer: A. The synthesis subagent's AgentDefinition.tools does not include Task, so it cannot spawn the fact-checker.
The Task tool is the mechanism for spawning subagents, and any agent that needs to invoke subagents must have Task in its allowed tools. The synthesis subagent is acting as a coordinator, but its scoped tools exclude Task.
What is the most effective way to make the model consistently see normalized values from these tools?
- Apply a PreToolUse hook for both tools to reject any call that may return non-normalized values.
- Require both MCP server owners to standardize their return schemas before the agent is deployed.
- Add system prompt instructions for date and status conversion so the model can translate values itself.
- Apply a PostToolUse hook for both tools to return updatedToolOutput with the normalized values.
Show answer
Answer: D. Apply a PostToolUse hook for both tools to return updatedToolOutput with the normalized values.
PostToolUse runs after a tool returns and can replace the tool output with updatedToolOutput before Claude processes it, giving deterministic normalization at the integration layer. System prompt instructions are probabilistic, and PreToolUse fires before the tool executes so it cannot transform the result.
Which decomposition best addresses this?
- A plan that maps checkout flow structure and high-impact failure areas, then creates a prioritized plan that adapts as dependencies emerge.
- A plan that increases parallel subagents for more query variations and asks each to return the top 10 relevant sources before synthesizing.
- A plan that processes failure categories in a fixed payment, inventory, shipping order, then combines reports without revisiting earlier categories.
- A plan that uses one document-analysis subagent to read every source in one context window and write the final root-cause report.
Show answer
Answer: A. A plan that maps checkout flow structure and high-impact failure areas, then creates a prioritized plan that adapts as dependencies emerge.
For open-ended investigation, the coordinator should first map the structure and high-impact areas, then generate subtasks that adapt to discovered dependencies. Broad parallel searches without structure create noise and miss relationships such as payment retries interacting with inventory locks. A fixed category pipeline or a single massive context window is also too rigid or too diluted to handle the discovery process.