Agentic Architecture & Orchestration

Question 1 · 1.1

A support agent sometimes stops after its first tool call, telling the customer “Let me look that up” and then nothing else. The loop exits whenever the response contains any text.

What is the most appropriate fix?

• ARaise the iteration limit from 5 to 20 so the loop can finish.
• BExit only when stop_reason is end_turn, and continue while it is tool_use.
• CTell the model in the system prompt never to write text before calling a tool.
• DExit when the text contains a phrase such as “Is there anything else?”

Question 2 · 1.1

A single response contains three tool_use blocks. What must the next request contain?

• AOne tool_result for whichever tool finished first.
• BThe assistant turn, then one user turn with three matching tool_result blocks.
• CThree separate requests, one for each tool result in turn.
• DOnly the tool outputs as plain user text; the ids are optional.

Question 3 · 1.1

A document-processing agent built on the Agent SDK is configured with max_turns=10. On a very long contract it stops, and the application shows an empty summary to the user.

What should the application do?

• AShow the user whatever text the last assistant message contained.
• BRemove max_turns entirely so that the agent can never be cut off again.
• CCheck ResultMessage.subtype; on error_max_turns, resume with a higher limit or report it.
• DAsk the model in the system prompt to finish within ten turns.

Question 4 · 1.2

A multi-agent research system is asked about the environmental impact of data centres. The report covers energy use thoroughly but says nothing about water consumption or e-waste. Each subagent completed its task without error.

Where is the fault most likely to be?

• AThe search subagent's tool descriptions, which limit what it can find.
• BThe synthesis subagent's context window, which truncated findings.
• CThe coordinator's decomposition, which scoped the topic too narrowly.
• DThe model's temperature setting, which made the plan less varied.

Question 5 · 1.2

Which TWO are benefits of routing all subagent communication through the coordinator? (Select 2.)

• AOne place to observe and log the system's behaviour.
• BLower latency than direct agent-to-agent messages.
• CConsistent error handling across all subagents.
• DSubagents share conversation history automatically.
• ESubagents no longer need their own tool restrictions.

Question 6 · 1.2

An internal research assistant runs search, analysis, synthesis and report-writing subagents for every request. Staff like the long reports, but finance flags that one-line questions such as “When was our Leeds office opened?” cost as much as full research tasks.

What is the best change?

• AHave the coordinator assess each query and spawn only the subagents it needs.
• BSwitch every subagent to the smallest available model to cut the cost.
• CCap every run at two subagents so no request can become expensive.
• DTell users to put simple questions to a different chat tool instead.

Question 7 · 1.2

A coordinator asks three subagents to “research the semiconductor shortage.” Two of them return nearly identical reports on current supply chains; the third covers the history of the crisis.

What change would most directly prevent the duplication?

• AGive the subagents a shared scratchpad so they can see each other.
• BRun the three subagents one after another instead of in parallel.
• CAdd a de-duplication step after synthesis to drop repeated material.
• DGive each subagent a distinct objective, boundaries and output format.

Question 8 · 1.3

A synthesis subagent produces a report that ignores most of the evidence gathered earlier. Logs show the coordinator spawned it with the prompt: “Write the final report from the research so far.”

What is the most likely cause?

• AThe synthesis subagent runs on a model that is too small for the job.
• BThe synthesis agent lacks the web search tool it needs to gather data.
• CIt started fresh, so it never received the research in its prompt.
• DSubagents must be resumed rather than spawned to see earlier work.

Question 9 · 1.3

A coordinator needs results from four independent search subagents. How should it spawn them to minimise total time?

• AEmit four spawn calls in a single coordinator response.
• BSpawn one subagent, wait for its result, then spawn the next.
• CGive one subagent all four topics in a single long prompt.
• DHave each subagent spawn the next one when it has finished.

Question 10 · 1.3

A team defines three subagents with clear descriptions and passes them in agents. The coordinator's allowed tools are Read, Grep and Glob. In testing, the coordinator never delegates and does all the work itself.

What should they check first?

• AWhether each subagent's model field names a supported model.
• BWhether the spawning tool is in the coordinator's allowed tools.
• CWhether the subagents' tools lists include Read and Grep.
• DWhether the coordinator's system prompt forbids doing work itself.

Question 11 · 1.3

Three analysis subagents each return findings with a URL and page number. The coordinator summarises their results in a paragraph and passes it to the synthesis subagent. The final report has almost no citations.

What is the best fix?

• ATell the synthesis subagent to search again for the missing sources.
• BAdd a citation-checking subagent after the report is written.
• CGive the synthesis subagent the same system prompt as the analysts.
• DPass the findings as structured claims, each with source and page.

Question 12 · 1.4

Policy requires identity verification before any refund. The system prompt says so in capital letters, yet an audit finds 0.4% of refunds were issued without verification.

What is the most effective change?

• AMove the instruction to the top of the system prompt and repeat it.
• BAdd few-shot examples that show verification before each refund.
• CSwitch to a more capable model that follows instructions better.
• DBlock process_refund until get_customer returns a verified id.

Question 13 · 1.4

An agent escalates to a human team who cannot see the conversation. What should the handoff contain?

• AThe full raw transcript, so nothing at all is lost in the handover.
• BA sentiment score and the customer's most recent message.
• CCustomer id, root cause, actions taken and a recommended step.
• DNothing — the human should ask the customer to repeat the issue.

Question 14 · 1.4

A bank is reviewing the rules in its account-servicing agent's system prompt.

Which TWO rules most need programmatic enforcement rather than prompt instructions? (Select 2.)

• AConfirm the caller's identity before showing any balance.
• BAddress the customer by first name where appropriate.
• CNever move more than the daily transfer limit automatically.
• DKeep replies under about 150 words unless asked for detail.
• EOffer a satisfaction survey link at the end of each chat.

Question 15 · 1.4

A customer writes: “My order is late, I was charged twice, and I need to change my delivery address.” The agent refunds the duplicate charge and closes the conversation. The other two issues are never addressed.

What design change best prevents this?

• ADecompose the message into items, handle each, then send one reply.
• BEscalate every message that mentions more than one problem to a human.
• CAsk the customer to send each problem in a separate conversation.
• DRaise max_tokens so the agent has room to write a longer reply.

Question 16 · 1.5

Refunds above $500 must never be issued automatically. Which implementation guarantees this?

• AA system prompt rule backed by three few-shot examples.
• BA PostToolUse hook that reverses any large refund.
• CA PreToolUse hook denying refunds over $500.
• DLowering the model's temperature for refund turns.

Question 17 · 1.5

Three MCP tools return dates as a Unix timestamp, an ISO 8601 string and a locale string. The agent occasionally misorders events. What is the best fix?

• AA PostToolUse hook that converts every date to ISO 8601.
• BExplain in the system prompt how to parse each date format.
• CA PreToolUse hook that blocks the tools returning timestamps.
• DAdd a fourth tool that compares dates on the model's behalf.

Question 18 · 1.5

A PreToolUse hook blocks deletions outside a project folder. It returns deny with no reason. Logs show the agent retrying the same deletion several times, then telling the user it “encountered an error”.

What change best improves the behaviour while keeping the block?

• ASwitch the hook to allow and log the deletions for later review.
• BAdd a reason that says why and what to do instead.
• CAdd a system prompt line telling Claude not to retry denied calls.
• DRemove the deletion tool from the agent's tool set completely.

Question 19 · 1.5

A CRM MCP tool returns full customer records, including card numbers. Policy says card numbers must never enter the model's context, but the agent needs the rest of the record.

Where should the masking happen?

• AIn the system prompt, telling Claude never to repeat card numbers.
• BIn a PreToolUse hook that denies every call to the CRM tool.
• CIn a PostToolUse hook returning updatedToolOutput with masking.
• DIn a separate subagent that reads the record and then summarises it.

Question 20 · 1.6

A CI review of a 40-file pull request reports shallow comments and misses a bug where a renamed field in one module was not updated in two others.

Which change addresses both problems?

• AUse a model with a larger context window and review all 40 files at once.
• BAsk the model in the prompt to be thorough and check every file carefully.
• CReview only the files with the most changed lines in this pull request.
• DReview each file separately, then run a cross-file integration pass.

Question 21 · 1.6

Which task is best suited to dynamic, adaptive decomposition rather than a fixed pipeline?

• AChecking every invoice for the same five required fields.
• BFinding why an unfamiliar service drops messages.
• CTranslating each release note into three languages.
• DFormatting a quarterly report to the house template.

Question 22 · 1.6

Which TWO are good reasons to implement a workflow as a fixed prompt chain? (Select 2.)

• AEvery input needs the same steps in the same order.
• BEach step can be tested and debugged on its own.
• CThe next step depends on what the previous one finds.
• DThe task is exploratory and its scope is unclear.
• EIt gives the lowest possible end-to-end latency.

Question 23 · 1.6

An agent is adding tests to a legacy codebase from a plan it wrote at the start. While testing the orders module, it discovers the module depends on an untested pricing library with inconsistent behaviour.

What should happen next?

• AContinue the original plan and note the pricing issue for the end.
• BUpdate the plan to characterise the pricing library before orders.
• CAbandon the plan and start again with a fresh map of the codebase.
• DSkip the orders module entirely because it has a risky dependency.

Question 24 · 1.7

Yesterday an agent analysed a service in a session. Overnight, a teammate refactored most of the service's modules. You want to continue the work today.

What is the most reliable approach?

• AResume yesterday's session and continue from where it stopped.
• BFork yesterday's session and continue the work in the fork.
• CStart fresh, seeded with a summary of findings that still hold.
• DResume the session and ask the agent to be extra careful.

Question 25 · 1.7

You have a session containing a detailed codebase analysis and want to compare two refactoring approaches without losing either. What should you do?

• AFork the session and pursue one approach in each branch.
• BRun both approaches one after the other in the same session.
• CStart two brand-new sessions and redo the analysis in each.
• DCompact the session first, then continue with both ideas.

Question 26 · 1.7

A developer forks a session to try an experimental refactor. The fork edits several source files. Later, resuming the original session, the agent finds the experimental changes in the code.

What explains this?

• AThe fork was created without fork_session, so it was not a fork.
• BForks copy conversation history, not the files on disk.
• CResuming a session always merges the history of its forks back in.
• DSession ids are shared between a fork and its original session.

Question 27 · 1.7

A CI job runs an agent on one ephemeral worker and saves only the session_id. A follow-up job on a different worker passes that id as resume and fails because no conversation is found.

What is the most robust fix?

• APass the same id with continue_conversation=True instead.
• BRetry the resume a few times in case the store is slow.
• CSave the results as job state and seed a fresh session.
• DPin both jobs to the same working directory path.