Deck 1 — Agentic Architecture & Orchestration (Domain 1, 27%)
60 cards. Largest deck for the largest domain. Cards 1–20 = the loop. 21–40 = multi-agent. 41–60 = enforcement, hooks, handoffs, sessions.
Use: cover the answer, say it out loud, then check. Move any card you miss to a separate pile and re-run that pile daily.
1
Q: What stop_reason value means the agentic loop must continue?
A: "tool_use".
2
Q: What stop_reason value means the loop terminates?
A: "end_turn".
3
Q: How do you add the model's turn to the message history in the loop? A: Append the assistant message verbatim — every content block, unmodified.
4
Q: An assistant turn contained three tool_use blocks. How many user messages carry the results?
A: One. All tool_result blocks for one assistant turn go in a single user message.
5
Q: What links a tool_result to the call it answers?
A: tool_use_id.
6
Q: Anti-pattern: terminating the loop by parsing phrases like "I'm done." Why wrong?
A: Natural-language phrasing varies. It produces both false positives and false negatives. Use stop_reason.
7
Q: Anti-pattern: using an iteration cap as the primary termination condition. Why wrong? A: It truncates legitimate multi-turn work. A cap is a safety net, not a termination condition.
8
Q: Anti-pattern: terminating when the assistant text is non-empty. Why wrong? A: Text and tool calls co-occur in a single turn. Non-empty text says nothing about whether tools were requested.
9
Q: Anti-pattern: terminating after the first tool result. Why wrong? A: Real tasks need several turns — a lookup often only enables the next call.
10
Q: When is a model-driven decision the right architecture? A: When inputs vary, the path can't be enumerated in advance, and judgement is required.
11
Q: When is a pre-configured decision tree the right architecture? A: When the path is fixed, must be auditable, and must be identical every time.
12
Q: tool_choice: "auto" — behaviour?
A: The model decides whether to use a tool. Default.
13
Q: tool_choice: "any" — behaviour?
A: The model must use some tool, its choice which.
14
Q: How do you force one specific tool?
A: tool_choice: {"type": "tool", "name": "extract_invoice"}.
15
Q: What is the loop's driving signal, in one phrase?
A: The API's structured stop_reason, never interpreted text.
16
Q: A tool errors mid-loop. Does the loop stop?
A: No. Return the error as a tool_result with isError: true and let the model adapt.
17
Q: Where does an iteration cap legitimately belong? A: As a runaway-loop backstop, set far above any expected real iteration count.
18
Q: Two tools were called in one turn and one failed. What do you send back?
A: Both results, in one user message — the success and the error, each with its tool_use_id.
19
Q: What breaks if you summarize or reformat the assistant turn before appending it?
A: The tool_use blocks and their IDs are lost, so results can't be matched. Append verbatim.
20
Q: Which SDK does the exam associate with building custom agents like the support agent? A: The Claude Agent SDK.
21
Q: In a coordinator–subagent system, which topology does the exam expect? A: Hub-and-spoke — all communication passes through the coordinator.
22
Q: Can subagents communicate directly with each other? A: No. Everything routes via the coordinator.
23
Q: What context does a subagent inherit from the coordinator? A: Nothing. Isolated context. Not the conversation, not prior findings, not the original question.
24
Q: Consequence of subagent context isolation for prompt writing? A: Every delegating prompt must state literally every identifier, path, constraint, and prior finding needed.
25
Q: Which tool does a coordinator use to delegate?
A: The Task tool.
26
Q: A coordinator does all the work itself instead of delegating. First thing to check?
A: Whether "Task" is in its allowedTools.
27
Q: How do you run subagents in parallel?
A: Emit multiple Task calls in a single coordinator response.
28
Q: Is there a parallel: true flag for subagents?
A: No. Non-existent option; a distractor.
29
Q: What three things does an AgentDefinition carry?
A: A description (coordinator uses it for routing), a prompt, and scoped tools.
30
Q: What is the description field of an AgentDefinition actually for?
A: The coordinator reads it to decide which subagent to route to.
31
Q: Symptom: the coordinator's subtasks all cluster into one or two dimensions of a multi-dimensional question. Name the bug. A: Overly narrow decomposition.
32
Q: Fix for narrow decomposition? A: Explicit scope partitioning across named dimensions, plus a coverage check that re-delegates for gaps.
33
Q: Narrow decomposition — which layer do you blame? A: The coordinator's decomposition. The subagents worked correctly within the scope they were given.
34
Q: Three subagents return overlapping, duplicated sources. Fix? A: Scope partitioning — disjoint scopes, each prompt stating what the others cover.
35
Q: Name the pattern: choosing which subagents to invoke based on what the task actually needs. A: Dynamic subagent selection.
36
Q: Name the pattern: checking coverage of returned results and re-delegating for what's missing. A: Iterative refinement (re-delegation).
37
Q: Goal-oriented prompts or procedural step lists for subagents? A: Goal-oriented — state the goal and the quality criteria, not the procedure.
38
Q: Why do procedural subagent prompts fail? A: They're brittle — they break when the situation deviates from the assumed steps, and they can't adapt.
39
Q: Should each subagent get the full tool set? A: No. Least privilege — only the tools its role needs. Also improves selection accuracy.
40
Q: Prompt chaining vs dynamic adaptive decomposition — which when? A: Chaining for a fixed sequence of known steps; dynamic decomposition when later steps depend on what earlier ones discover.
41
Q: Deterministic vs probabilistic — classify hooks and prompt instructions. A: Hooks = deterministic. Prompt instructions = probabilistic.
42
Q: The exam's phrase for why prompt instructions aren't enough for financial actions? A: "Prompt instructions have a non-zero failure rate."
43
Q: When must you use programmatic enforcement rather than prompt guidance? A: When errors have financial, compliance, or irreversible consequences.
44
Q: Which hook blocks or redirects a tool call?
A: PreToolUse.
45
Q: Which hook transforms a tool result?
A: PostToolUse.
46
Q: Requirement: process_refund must never run before a verified customer ID exists. Implementation?
A: A PreToolUse gate that blocks the call when no verified ID is in session state.
47
Q: Requirement: refunds over $500 must go to a human. Implementation?
A: A PreToolUse interception that blocks the refund and redirects to escalate_to_human.
48
Q: Tools return timestamps as Unix epoch, ISO 8601, and numeric status codes. Fix?
A: A PostToolUse hook that normalizes all of them before they reach the model.
49
Q: Why not just describe the timestamp formats in the system prompt? A: It's probabilistic. The model will misread formats some fraction of the time. Normalization is deterministic.
50
Q: What belongs in a structured handoff to a human in the support scenario? A: Customer ID · root cause · refund amount · recommended action.
51
Q: Should the handoff include the full conversation transcript? A: No — a structured summary. The transcript makes the human re-derive what the agent already knows.
52
Q: Which three boundaries use the same structured-summary principle? A: Agent→human, agent→agent, tool→model.
53
Q: Symptom: a multi-file review misses issues it catches when reviewing one file. Name it. A: Attention dilution.
54
Q: Fix for attention dilution? A: Per-file passes plus a separate cross-file integration pass.
55
Q: Does a larger context window fix attention dilution? A: No. It's attention distribution, not capacity.
56
Q: What does the cross-file integration pass catch that per-file passes cannot? A: Cross-boundary problems — a changed signature with unchanged callers, contract mismatches across modules.
57
Q: Resuming a named session — flag?
A: --resume <session-name>.
58
Q: Branching a session to try a variant without losing the original?
A: fork_session.
59
Q: Prior session's tool results are now stale. Resume or start fresh? A: Fresh session plus a structured summary. Resuming would carry stale results forward.
60
Q: Same session, context filling, history still valid?
A: /compact.