Question Bank — Domain 1: Agentic Architecture & Orchestration

30 items. The largest domain (27%, ~16 items on the exam).

Every distractor has a rationale. Read them — the rationales are where the learning is, and the exam's wrong answers recur across items.

How to use: work a block of 10 under time (90 s/item), then read all ten answers. Log every miss against the Core Reflex it violates.

Block A — The agentic loop (items 1–8)

1

A support agent built on the Agent SDK terminates prematurely on about 15% of conversations, usually right after its first tool call. Review of the implementation shows the loop exits when the assistant response contains any text content. What is the correct termination condition?

A. Exit when stop_reason is not "tool_use".
B. Exit when the assistant response contains no tool_use blocks and the text is non-empty.
C. Exit after a maximum of 10 iterations.
D. Exit when the model's text output includes a completion phrase such as "I've resolved this."

Answer

stop_reason == "tool_use" means the model is requesting tool execution and the loop must continue. Any other value — normally "end_turn" — means the turn is complete. This is the API's structured signal and the only reliable condition.

B — Closer, but it adds a text check that isn't needed and can be wrong: a final turn may legitimately have empty text (e.g. content that's entirely a different block type). The stop_reason check alone is both necessary and sufficient. The stated bug in the stem is text-checking.
C — An iteration cap is a runaway-loop safety net, never a termination condition. It truncates legitimate multi-turn work — exactly the 15% failure being reported, in a different form.
D — Parsing natural-language termination signals. Phrasing varies; you get both premature exits and infinite loops.

Reflex: the loop is driven by structured signals, never interpreted text.

2

An agent makes three tool calls in a single assistant turn. How should the results be returned to the model?

A. As three separate user messages, one per result, in call order.
B. As a single user message containing three tool_result blocks, each with its tool_use_id.
C. As a single user message with the three results concatenated into one text block.
D. As three separate assistant messages, since the tools were called by the assistant.

Answer

All tool_result blocks corresponding to one assistant turn go in one user message, each keyed by the tool_use_id of the call it answers.

A — Breaks the turn structure. The API expects one user message answering one assistant turn.
C — Loses the tool_use_id correlation, so the model can't tell which result belongs to which call. With three parallel calls this is actively ambiguous.
D — Tool results are user-role content. The assistant role is for the model's own output.

3

Which of the following is a legitimate use of an iteration cap in an agentic loop? (Select one.)

A. As the primary termination condition, tuned to the typical task length.
B. As a backstop against runaway loops, set well above any expected iteration count.
C. To force the agent to summarize its progress at regular intervals.
D. Iteration caps should never be used.

Answer

A cap is a safety mechanism. Set far above realistic task length, it never fires on legitimate work but bounds a pathological loop.

A — The named anti-pattern. Tuned to "typical" length, it truncates every above-average task.
C — Conflates two mechanisms. Progress summarization is a separate design choice, not a function of the cap.
D — Overcorrection. Unbounded loops are a real operational hazard.

4

A team is designing the triage step of a customer support agent. Incoming issues vary widely and cannot be enumerated in advance. Which architecture fits?

A. A pre-configured decision tree covering the known issue categories, with a fallback branch.
B. Model-driven decision-making, with the model selecting tools based on the conversation.
C. A classifier model that routes to one of five fixed handlers.
D. A rules engine maintained by the support operations team.

Answer

Unbounded input variety and required judgement are exactly the conditions where model-driven decisions beat enumeration.

A — A decision tree is right when the path is fixed and must be auditable. Here it can't be enumerated, and the "fallback branch" would absorb most traffic.
C — Same problem with a different implementation: five fixed handlers is enumeration.
D — Maintaining rules for unbounded variety is unsustainable, and out of step with the technologies under test.

⚠️ Contrast with an item describing a compliance-mandated approval sequence — there, the pre-configured path (or a hook) is correct.

5

A refund tool occasionally executes before customer verification, despite a system prompt instruction stating that get_customer must always be called first. Refunds are financially material. What is the most effective fix?

A. Strengthen the instruction with emphatic wording and place it at the start of the system prompt.
B. Add a PreToolUse hook that blocks process_refund unless a verified customer ID exists in session state.
C. Add a few-shot example showing the correct call ordering.
D. Combine the two tools so verification happens inside process_refund.

Answer

Prompt instructions have a non-zero failure rate. When the consequence is financial, the requirement must be enforced deterministically. A PreToolUse gate blocks the call regardless of how the model was prompted or persuaded.

A — Firmer wording on a probabilistic mechanism. Reduces the rate; doesn't eliminate it.
C — Few-shot examples improve consistency but remain guidance, not enforcement.
D — Superficially appealing but wrong: it hides the verification step from the model's reasoning, removes the ability to handle ambiguous matches conversationally, and conflates two distinct capabilities. The gate is the targeted fix.

Reflex 1.

6

An agent calls four tools whose outputs represent timestamps inconsistently: one returns Unix epoch integers, one returns ISO 8601 strings, and one returns numeric status codes alongside dates. The agent makes date-arithmetic errors. Best fix?

A. Document each tool's format in its description so the model knows what to expect.
B. Add a PostToolUse hook that normalizes timestamps and status codes to a single representation before results reach the model.
C. Instruct the model in the system prompt to convert all timestamps to ISO 8601 before reasoning about them.
D. Use a model with stronger arithmetic capability.

Answer

Format normalization is deterministic work. Doing it in a PostToolUse hook means the model never sees heterogeneous formats and cannot misread them.

A — Better descriptions help selection, not arithmetic. The model still has to do the conversion, probabilistically, on every call.
C — Same problem: asks the model to do reliably what code can do certainly.
D — Model capability isn't the constraint; format heterogeneity is.

Reflex 1, applied to PostToolUse.

7

Which statement about the assistant message in an agentic loop is correct?

A. It should be appended verbatim, with all content blocks unmodified.
B. It should be summarized to conserve context before appending.
C. Only the tool_use blocks need to be appended.
D. It should be appended after the tool results, to preserve chronology.

Answer

Verbatim. The tool_use blocks and their IDs must survive intact so results can be correlated, and the model's own reasoning text is part of the turn.

B — Summarizing destroys the tool_use IDs and the turn's structure. Context management happens by trimming tool result payloads, not by mangling assistant turns.
C — Discards the model's reasoning text, which is part of its own turn.
D — Order matters: the assistant turn precedes the user message carrying its results.

8

A tool call fails with a transient network timeout mid-loop. What should the loop do?

A. Terminate and surface the error to the user.
B. Return the failure as a tool_result with isError: true and continue the loop.
C. Silently retry the tool up to three times, then return an empty result.
D. Skip the result and let the model proceed without it.

Answer

Errors are information. Returned as a tool_result with isError: true and a category, the model can retry, choose a different tool, or explain the limitation. A transient category tells it retrying is worthwhile.

A — Terminating the whole flow on one recoverable failure discards all progress. Named anti-pattern.
C — Retrying is reasonable; returning an empty result on exhaustion is not. An access or transport failure reported as "nothing found" is silently wrong (Reflex 6).
D — The model then reasons over a gap it doesn't know exists.

Block B — Multi-agent orchestration (items 9–20)

9

A research coordinator delegates to four subagents. Reviewing the logs, an architect sees that for a question spanning product, pricing, distribution, regulation, and financials, the coordinator's four subtasks all concerned product features and pricing. The subagents returned accurate, well-sourced results within those areas. What is the root cause?

A. The subagents' tool access is too narrow.
B. The coordinator's task decomposition is overly narrow.
C. The synthesis agent is dropping dimensions during merging.
D. The subagents need larger context windows to cover more ground.

Answer

The stem tells you the subagents performed accurately within the scope they were given. The dimensions that are missing were never delegated. That's a decomposition failure at the coordinator.

A — Tool access would show up as failed or empty retrievals, not as missing dimensions.
C — Synthesis can only drop what it received. Nothing about regulation ever reached it.
D — The classic capacity distractor. Context size has nothing to do with which subtasks were created.

Reflex 9: blame the layer the evidence implicates.

10

Fix for the situation in item 9? (Select one.)

A. Instruct the coordinator to explicitly partition the question across named dimensions and to check coverage against them, re-delegating for any gaps.
B. Increase the number of subagents from four to eight.
C. Give the coordinator a procedural checklist of research steps to follow in order.
D. Have each subagent research the entire question independently and merge the results.

Answer

Scope partitioning across explicitly named dimensions plus an iterative-refinement coverage check. That addresses both halves: initial breadth and detection of residual gaps.

B — More subagents with the same narrow decomposition produce more narrow subtasks.
C — Procedural prompts are brittle and don't generalize to the next question. The guide prefers goal-oriented prompts with quality criteria.
D — Deliberate duplication: four agents doing the same broad work, wasting effort and producing overlapping sources. This is what scope partitioning exists to prevent.

11

In a coordinator–subagent architecture, how should two subagents share intermediate findings?

A. Directly, via a shared message channel.
B. Through the coordinator, which passes relevant findings into subsequent delegating prompts.
C. Through a shared scratchpad file both can write to.
D. They cannot share findings; each must work entirely independently.

Answer

Hub-and-spoke: all inter-subagent communication routes through the coordinator, which owns the composition of context for each delegation.

A — Not the architecture the exam describes, and it destroys the coordinator's ability to reason about overall state.
C — Scratchpad files are a context-management mechanism for a single agent's long task (Domain 5), not the inter-agent communication channel. Tempting because it's a real pattern from a different context.
D — Too absolute. Findings do flow — via the coordinator.

12

What context does a subagent receive when a coordinator delegates to it?

A. The full conversation history plus the coordinator's system prompt.
B. Only what the delegating prompt explicitly contains.
C. The conversation history but not prior subagent results.
D. A compressed summary of the session, generated automatically.

Answer

Subagents have isolated context and inherit nothing. Everything they need must be in the delegating prompt.

A, C, D — All assume some form of inheritance. None occurs. C is the most seductive because partial inheritance sounds plausible; it isn't the model.

13

A coordinator's delegating prompt reads: "Research the regulatory position for the company we discussed, over the timeframe mentioned earlier." The subagent returns irrelevant results. Why?

A. The subagent lacks web search tools.
B. The prompt relies on context the subagent does not have.
C. The prompt is too long and the key instruction is lost in the middle.
D. The subagent's model is too small for regulatory analysis.

Answer

"The company we discussed" and "the timeframe mentioned earlier" refer to the coordinator's context, which the subagent never received. It has no company name and no dates.

A — Tool absence produces failures or empty results, not irrelevant ones.
C — The prompt is two clauses long. Lost-in-the-middle needs length.
D — Nothing suggests a capability gap; the input was incomplete.

Reflex 5.

14

A coordinator agent consistently performs research itself rather than delegating, despite a prompt instructing it to use subagents. What should be checked first?

A. Whether "Task" is included in the coordinator's allowedTools.
B. Whether the subagents' AgentDefinition prompts are sufficiently detailed.
C. Whether the coordinator's model supports multi-agent orchestration.
D. Whether parallel: true is set in the coordinator's configuration.

Answer

Without "Task" in allowedTools the coordinator has no delegation mechanism available, so it does the work itself. The behaviour matches exactly.

B — Would produce poor subagent output, not zero delegation.
C — Orchestration isn't a model capability flag.
D — parallel: true does not exist. Parallelism comes from multiple Task calls in one response.

15

Four research subtasks are fully independent. How does the coordinator run them concurrently?

A. Set parallel: true in the coordinator's configuration.
B. Emit four Task tool calls in a single assistant response.
C. Emit one Task call per turn and rely on the SDK to pipeline them.
D. Define the four subagents with concurrency: 4 in their AgentDefinition.

Answer

Multiple Task calls in one response is the parallelism mechanism.

A, D — Non-existent options.
C — Sequential by construction. Each turn waits for the previous result.

16

What is the purpose of the description field in an AgentDefinition?

A. Documentation for human maintainers of the agent configuration.
B. The coordinator uses it to decide which subagent to route a task to.
C. It becomes the subagent's system prompt.
D. It is shown to the end user when the subagent is invoked.

Answer

Routing. The coordinator reads subagent descriptions to select one, in the same way the model reads tool descriptions to select a tool.

A — It has a functional role, not just a documentary one.
C — That's the separate prompt field.
D — Not a user-facing string.

🧠 Note the parallel with Domain 2: descriptions drive selection, at both the tool and the subagent level.

17

Three research subagents are given the same broad instruction: "research this company's competitive position." The returned results overlap heavily and cite many of the same sources. Which pattern addresses this?

A. Scope partitioning — disjoint scopes, each prompt stating what the other agents cover.
B. Dynamic subagent selection based on the question type.
C. Iterative refinement with a coverage check.
D. Reducing to a single subagent to eliminate duplication.

Answer

Duplication is precisely what scope partitioning prevents. Telling each agent what the others are handling further suppresses overlap.

B — Addresses which subagents to invoke, not how to divide work between the ones you've chosen.
C — Detects and fills gaps. Useful, but the stem's problem is overlap, not omission. A plausible half-right answer.
D — Eliminates duplication and parallelism together. Over-correction.

18

Which subagent prompt style does the guide favour?

A. A numbered procedure the subagent follows in order.
B. A statement of the goal and the quality criteria for a good result.
C. A minimal one-line task with no constraints, to maximize flexibility.
D. A full copy of the coordinator's system prompt plus the task.

Answer

Goal-oriented prompts with explicit quality criteria. The subagent can adapt its approach while still being held to a standard.

A — Procedural prompts are brittle; they break on any deviation from assumed conditions.
C — Underspecified. Subagents inherit nothing, so a one-liner gives them nothing to work with.
D — Bloats context with irrelevant instructions and doesn't state the task's own criteria.

19

A research pipeline has a fixed sequence: retrieve documents → extract claims → verify claims → draft report. Which approach fits?

A. Prompt chaining.
B. Dynamic adaptive decomposition.
C. Hub-and-spoke coordination with dynamic subagent selection.
D. A single agent with all tools and no decomposition.

Answer

The steps are known and fixed. Chaining is the simple, appropriate structure.

B — For when later steps depend on what earlier ones discover. Over-engineering here.
C — Adds routing machinery for a sequence that never varies.
D — Loses the staging benefits (isolated context per stage, targeted prompts) for no gain.

20

A coordinator receives structured findings from three subagents and needs to determine whether a fourth research dimension is required. Which pattern is this?

A. Scope partitioning.
B. Iterative refinement with re-delegation.
C. Prompt chaining.
D. Attention dilution mitigation.

Answer

Assessing what's returned and re-delegating for what's missing is iterative refinement.

A — The upfront division of work, not the post-hoc gap check.
C — Chaining is a fixed sequence; this is a conditional decision based on results.
D — A different problem (per-file passes for multi-file work).

Block C — Enforcement, handoffs, sessions (items 21–30)

21

Policy requires refunds above $500 to be approved by a human. Which implementation satisfies this most reliably?

A. A system prompt instruction that refunds above $500 must be escalated.
B. A PreToolUse hook that blocks process_refund when the amount exceeds $500 and redirects to escalate_to_human.
C. A PostToolUse hook that logs refunds above $500 for later audit.
D. Removing process_refund from the agent's tools and requiring all refunds to be escalated.

Answer

Deterministic interception at the boundary, with a redirect so the workflow continues correctly rather than just failing.

A — Probabilistic, for a compliance requirement.
C — Detects violations after the money has moved. Audit is not prevention.
D — Satisfies the letter of the policy by destroying the capability. Every sub-$500 refund now needs a human, defeating the first-contact-resolution goal. Over-correction.

22

When a support agent escalates to a human, what should the handoff contain?

A. The complete conversation transcript so the human has full context.
B. A structured summary: customer ID, root cause, refund amount, and recommended action.
C. A one-line reason for the escalation.
D. The transcript plus the agent's confidence score for each proposed action.

Answer

Structured, minimal, complete. The human gets the conclusions the agent already reached rather than re-deriving them.

A — Forces the human to re-read and re-analyse everything, which is the cost the agent existed to remove.
C — Too little. The human has to start over.
D — Adds transcript bloat plus uncalibrated confidence scores. Two anti-patterns in one option.

23

A code review agent operating across 40 changed files reports fewer issues per file than when reviewing the same files individually. What is happening, and what is the fix? (Select one.)

A. Attention dilution; run per-file review passes plus a separate cross-file integration pass.
B. Context window exhaustion; switch to a model with a larger context window.
C. The prompt is too vague; add explicit severity criteria.
D. The agent needs extended thinking enabled for large reviews.

Answer

Reduced per-file sensitivity at scale is attention dilution. Per-file passes restore local focus; the integration pass covers what per-file passes structurally cannot see.

B — The capacity distractor. Dilution is an attention-distribution problem; a bigger window doesn't concentrate attention.
C — A real technique for a different problem (false positives / inconsistent severity). Nothing in the stem suggests vagueness — the same prompt worked per file.
D — More reasoning over diluted attention.

Reflex 2.

24

Why is a separate cross-file integration pass required in addition to per-file passes?

A. To reduce cost by batching the per-file findings.
B. To catch issues that span file boundaries, such as a changed function signature with unchanged callers.
C. To deduplicate findings reported by multiple per-file passes.
D. To give the model a second chance at the files it reviewed first.

Answer

A per-file pass sees one file. Contract mismatches across modules are invisible to it by construction.

A — Not a cost mechanism.
C — Deduplication is a real housekeeping task but not the reason for the pass.
D — The pass has a distinct purpose, not a repeat one.

25

A developer returns to a multi-day refactoring task. The workspace is unchanged and yesterday's session recorded useful decisions. What is the appropriate mechanism?

A. --resume <session-name>.
B. fork_session.
C. A fresh session with a structured summary of the prior work.
D. /compact on the prior session.

Answer

Continuing a task where the history is still valid is exactly what resume is for.

B — For branching to explore a variant while keeping the original.
C — Right when the prior session's tool results are stale — files have changed since. The stem says the workspace is unchanged.
D — Compaction addresses a filling context within an active session, not returning to an ended one.

26

An agent resumed a session from three days ago and began proposing edits based on file contents that had since changed. What should have been done?

A. Run /compact before resuming.
B. Start a fresh session and provide a structured summary of the prior decisions.
C. Use fork_session to branch from the old session.
D. Re-run the same session with a larger context window.

Answer

Stale tool results are the failure. A fresh session re-reads current state; the structured summary carries forward the decisions worth keeping.

A — Compaction compresses the stale results; it doesn't refresh them.
C — Forking inherits the same stale results.
D — Capacity distractor; staleness isn't a size problem.

27

Which pairs of mechanism and purpose are correct? (Select two.)

A. PreToolUse — blocking or redirecting a tool call before it executes.
B. PostToolUse — normalizing or trimming a tool result before it reaches the model.
C. PreToolUse — summarizing the conversation before the next model call.
D. PostToolUse — enforcing that a prerequisite tool was called first.

Answer

A and B.

Pre = intercept the call. Post = transform the result.

C — Not what PreToolUse does; conversation compaction is a separate mechanism.
D — Wrong side of the boundary: by the time PostToolUse runs, the call has already executed. Prerequisite enforcement must be PreToolUse.

28

A support agent's refund tool is gated on a verified customer ID. A customer name matches three accounts. What should the agent do?

A. Select the account with the most recent activity and proceed.
B. Ask the customer for an additional identifying detail such as an email address or ZIP code.
C. Escalate to a human immediately, since verification failed.
D. Proceed with the refund and flag the ambiguity for audit.

Answer

Verification means unambiguous identification. The agent asks for one more identifier — a cheap, in-conversation resolution.

A — Guessing. The downstream action is an irreversible refund against a specific account.
C — Premature. Ambiguity the agent can resolve conversationally is not an escalation trigger; escalating here damages first-contact resolution.
D — Takes the irreversible action first and documents the risk afterwards.

29

Which of the following would be correctly implemented as a pre-configured decision path rather than left to model judgement? (Select one.)

A. Choosing which tool to use for an ambiguous customer question.
B. A compliance-mandated approval sequence that must execute identically every time and be auditable.
C. Deciding how to phrase an apology to a frustrated customer.
D. Determining which of five research dimensions still needs coverage.

Answer

Fixed, auditable, invariant — the defining conditions for pre-configuration.

A, C, D — All require judgement over varying inputs. D in particular is the iterative-refinement pattern, which is model-driven by design.

30

A coordinator delegates document analysis to a subagent, which returns four paragraphs of reasoning followed by its findings. Over a long research run the coordinator's answers become vague. Which two changes help most? (Select two.)

A. Specify a structured return format for subagent results — findings only, as {claim, source, date} records.
B. Increase the coordinator's context window.
C. Have the coordinator maintain a scratchpad file of accumulated findings.
D. Have subagents communicate findings directly to the synthesis agent, bypassing the coordinator.

Answer

A and C.

A stops the coordinator's context filling with reasoning it doesn't need and keeps findings positionally prominent. C makes accumulated findings survive independently of the context that collected them.

B — The capacity distractor. Verbose returns fill any window, and the findings still land in the middle.
D — Breaks hub-and-spoke. Subagents don't communicate directly.

Scoring

Score	Read
26–30	Strong. Move to Domain 2.
21–25	Solid. Re-read Chapters 03 and 04 on the misses.
16–20	Re-read Chapters 02–04 fully, then re-run this bank.
< 16	Domain 1 is 27% of the exam. Re-study before continuing.

Log each miss against a Core Reflex. If three or more misses trace to the same reflex, that reflex is your highest-value study target.