# Applied AI — Advanced Level — Session 4 # Quiz: Multi-agent architecture **Instructor:** Yann Isola **10 multiple choice questions — only one correct answer per question** **Recommended passing threshold: 8/10 (certification level)** --- **Q1.** A workflow is divided into three agents: the first produces a plan, the second executes each step, the third checks the result against the success criteria. What decomposition strategy is this? - A) Horizontal — each agent has its domain - B) Vertical — division by depth into stages of different natures - C) Recursive — each agent decomposes its own subtask - D) Federated — agents vote on the final result **Answer: B.** Planning → execution → verification is the canonical motif of vertical decomposition: *sequential* steps of *different natures*. The horizontal (A) cuts into parallel domains; recursive (C) involves tree delegation. --- **Q2.** What is the main advantage of horizontal decomposition over vertical? - A) It eliminates the need for an orchestrator - B) It guarantees the absence of contradictions between agents - C) Parallelism: latency tends towards that of the slowest agent, not towards the sum of the steps - D) It always consumes fewer tokens in total **Answer: C.** Independent domain agents work in parallel. It always requires an orchestrator (A false), on the contrary creates a risk of inter-domain inconsistencies to be reconciled (B false), and often consumes *more* tokens in total despite the reduced latency (D false). --- **Q3.** In a recursive architecture, what safeguard is essential to prevent an agent from creating an infinite chain of delegations? - A) A correlation ID propagated at each level - B) A maximum delegation depth (`max_depth`) and a legacy token budget per level - C) A circuit breaker on each external tool - D) A queue of dead letters per sub-agent **Answer: B.** The depth limit and the inherited budget structurally bound the recursion. The correlation ID (A) serves observability but does not prevent anything; C and D are error handling, not recursion control. --- **Q4.** A single, homogeneous task, fitting comfortably in the context window with satisfactory output quality, should be executed 1000 times per day. What is the correct architectural decision? - A) Horizontal decomposition: one agent per 100 executions - B) Vertical decomposition to add a verifier by principle - C) No decomposition of the task: the volume is processed by parallelizing instances of the same agent - D) Recursive decomposition to absorb the load **Answer: C.** None of the three criteria (complexity, token budget, specialization) justifies the decomposition. Volume is a problem of *instance scaling*, not of *task decomposition* — classic confusion tested in certification. --- **Q5.** A support agent responds with great confidence to a question about international taxation — a regulated area that he does not cover — and the answer is invented. Which escalation trigger should have occurred first?- A) Confidence below the threshold, because the agent should have doubted - B) Out-of-scope detection, because the self-reported confidence score of a guaranteed hallucination is precisely high - C) The SLA timeout - D) The circuit breaker of the tax tool **Answer: B.** The catch: a confident hallucination does not trigger the self-assessed confidence threshold — this is exactly why out-of-perimeter detection (upstream classifier + prompt, defense in depth) is an independent trigger. C and D are irrelevant to the case. --- **Q6.** An agent prepares funds withdrawal instructions. The policy requires that a human approve each instruction before execution, but the agent must not be stuck waiting. Which climbing pattern corresponds? - A) Pause-and-ask: the agent suspends its execution until the response - B) Queue-for-review: the agent finishes its preparation, the result awaits approval in a queue, the agent is released - C) Fallback-to-human: the agent transfers the entire task to the human - D) Fail-operational: the default option applies to the timeout **Answer: B.** The agent completes his work (preparation) and the deliverable goes to the validation queue before execution — this is the canonical case of queue-for-review. Pause-and-ask (A) would block the agent mid-task; fallback (C) would forego the preparation work; D is a timeout strategy, not an escalation pattern. --- **Q7.** What is the golden rule for a successful fallback-to-human? - A) Only transfer the initial request so as not to bias the human - B) Transfer the request, the history of the agent's actions, the hypotheses tested and the precise reason for escalation — humans never start from scratch - C) Delete the agent context for confidentiality reasons - D) Wait for the end of the SLA before transferring **Answer: B.** A transfer without context destroys the value of the work already accomplished and lengthens the processing. A and C provide this context; D confuses transfer and timeout. --- **Q8.** Why is tool idempotency a prerequisite for automatic retry? - A) Because it reduces the latency of tool calls - B) Because a timeout is ambiguous: the first attempt could have succeeded on the server side, and replaying a non-idempotent tool risks duplicating the effect (e.g.: double billing) - C) Because idempotent tools cannot fail - D) Because idempotence makes exponential backoff useless **Answer: B.** The timeout does not indicate whether the operation was successful; without idempotence (typically via deduplicated idempotence key on the server side), retry creates duplicates. A, C and D are unfounded — idempotence guarantees neither performance nor infallibility, and backoff remains necessary. --- **Q9.** A circuit breaker has been in the OPEN state for longer than its cool down delay. A new call comes. What is happening? - A) The call is rejected immediately, like all calls in open state - B) The circuit goes directly back to CLOSED and all calls resume normally - C) The circuit goes to SEMI-OPEN: this test call goes through; a success closes the circuit, a failure reopens it - D) The call is placed in a dead letter queue **Answer: C.** After the cooldown, the half-open state lets a test call pass which decides what happens next: success → CLOSED, failure → OPEN.B would skip the cautious verification step; Describes the open state *before* timeout; D concerns definitively failed tasks. --- **Q10.** A pipeline's enrichment agent is down. The system delivers the report with the four other sections and the mention “enrichment section unavailable (service down)”. What pattern is this, and what absolute rule illustrates the explicit mention? - A) Dead letter tail; replay rule after correction - B) Partial rollback; reverse order compensation rule - C) Graceful degradation; transparency rule — never present a degraded result as complete - D) Circuit breaker; fail fast rule **Answer: C.** Reducing capacity (4 out of 5 prongs) rather than failing completely = graceful degradation; the explicit warning applies the transparency rule. The DLQ(A) would store the failed subtask in parallel but does not describe partial delivery; B and D are other bosses in the chapter. --- ## Quick correction grid | Q1 | Q2 | Q3 | Q4 | Q5 | Q6 | Q7 | Q8 | Q9 | Q10 | |----|----|----|----|----|----|----|----|----|-----| | B | C | B | C | B | B | B | B | C | C | **Analysis of common errors:** - Q4 and Q5 are the discriminators of the certification level: volume ≠ decomposition, and self-declared confidence ≠ perimeter detection. - Q6 vs Q7: check that the distinction between the three climbing patterns is acquired; otherwise, return to §2.3 of the guide with a chronological example. - Q9: circuit breaker transitions are systematically evaluated during the exam — use the simulator on the web page in the event of collective failure.