Real Interview Questions for the Agentic AI Engineer at Amazon
New
Practice MCQs
Answers count toward the Agentic AI Engineer and overall leaderboards; your latest answer to each question is the one that counts.
MCQ-1A returns agent's new planner prompt lifts the offline benchmark from 91% to 94% on 500 replayed tickets from last quarter, which carry human labels and recorded tool responses; intents added since are absent. At 10% traffic the escalation guardrail rose from 4.2% to 9.8% over three days, with tool errors and p95 flat. Those traces hold full tool calls but no resolution labels, and the planner runs on a vendor endpoint the team cannot fine-tune. What should gate the next prompt iteration?⌄
MCQ-2An ops optimizer runs 40 planner agents on a Postgres `inventory` table. Each reads a SKU's `on_hand` and `version`, plans for 2–4 s, then writes `UPDATE inventory SET on_hand = :new, version = version + 1 WHERE sku = :sku AND version = :v`. All 40 hit one SKU within 200 ms: 3 writes land, 37 get 0 rows and re-plan at once, p99 goes 90 ms to 6 s. Moves are independent decrements guarded by stock, and any agent may take any SKU. Which change removes the storm and keeps stock right?⌄
MCQ-3A refund agent on a durable workflow engine checkpoints after every activity and keeps a full event history. It has issued 3 refunds of $40 each under the old $50 auto-approval rule and has 12 queued when the customer tightens the rule to $25 and confirms the 3 issued refunds must stand. Which change re-plans the queued work without resetting the durable state?⌄
MCQ-4A returns agent compacts at turn 30, keeping 10 turns verbatim and a 7B summary that drops `issue_refund` payloads; sessions run 60 turns of ~350 tokens on an 8K cap. Of 1,900 refunds, 14 were issued twice, all after a compaction: 9 first went through in an earlier session, 5 followed a payments 504 that posted anyway. Sessions idle up to 45 min between turns, and the vector store's recall@5 on exact `order_id` queries is 0.71. Where should the refund record live, and when read?⌄
MCQ-5A returns agent moves from 5×H100 80GB to one H100 80GB and must still serve 1.2 tickets/s at p95 ≤ 4.0 s. A ticket is 800 output tokens across a 70B planner, a 7B extractor and a 70B writer; decode measured 42 ms/token for the 70B BF16, and 71% of tickets replay one of 60 templates. The 400-ticket eval floor is 0.93: 70B BF16 0.96, 70B INT4 0.95, distilled 7B 0.94. Which plan fits the card and holds both bounds?⌄
MCQ-6A support agent's turn is 1.0 s of planning, three reads of one ticket record at 150 ms p95 each (issuable as a single 200 ms call), and one irreversible write (a refund or an email) at 300 ms, drawn from a fixed six-tool set. The supervisor costs 700 ms per call it scores and catches 96% of the red-team set; the distilled model costs 150 ms and catches 71%. The release gate is 90% recall on every call in the turn and the turn SLO is p95 3.0 s. Which design clears both?⌄
MCQ-7An agent calls `issue_refund` with `order_id` matching `^ord_[0-9]{8}$`, `amount_cents` an int 1–50000 and `reason` an enum. Constrained decoding from the schema file is on; the gateway validates that schema. The grammar blocked 340 calls with a string `amount_cents`, the gateway rejected 212 outgoing calls for a missing `reason`, yet 6% send `ord_99999999` with `amount_cents: 49999` on orders of $12; totals run $12–$480 and the API has no reversal endpoint. Which account of the two is right?⌄
MCQ-8A RAG service serves 10,000 QPS under a 200 ms p99 budget for embedding plus retrieval; generation streams after. Measured p99: embedding 45 ms, HNSW search over 50M chunks 120 ms, cross-encoder rerank of the top 100 at 60 ms. 225 ms. 62% of queries repeat one from the last 10 minutes. An offline eval pins recall@10 at 0.95 as a hard floor and shows reranking the top 20 instead of 100 leaves the final top-3 unchanged in 99.6% of cases. Which change brings p99 under 200 ms?⌄
MCQ-9A LangGraph supervisor routes to `researcher` until the graph's `messages` channel holds a `ToolMessage` from the search tool, then to `writer`, then END; `messages` is `Annotated[list, add_messages]`. The researcher wraps a prebuilt ReAct agent and returns `{"messages": [result["messages"][-1]]}`. Runs hit `GraphRecursionError` at 25 steps with 25 messages, all `AIMessage`, and `writer` never runs. Which change makes `writer` run only after search results exist?⌄
MCQ-10A retrieval service keeps 2 billion 1024-dim float32 embeddings on 256 GB nodes; exact vectors sit on local NVMe at 45 µs p99 per reranked candidate, and PQ search at nprobe=32 costs 21 ms p99. A 50M-vector pilot at nprobe=32 with a 200-candidate exact rerank measured recall@10 of 0.94 for 64-byte codes, 0.88 for 32-byte codes and 0.62 for 8-byte codes; the floor is 0.90 and search plus rerank must stay under 60 ms p99. Which configuration meets both limits at 2 billion?⌄
MCQ-11A code agent's runner starts a fresh container per attempt from the same image with `--network none`, a non-root user, the repo read-only, `PYTHONPATH=/sandbox/lib`, `HOME=/sandbox/home`, and one named volume mounted at `/sandbox`. After a failed attempt the next patch passes: a `sitecustomize.py` in `/sandbox/lib` and a patched `pytest` in `/sandbox/home/.local` make the failing test skip, and the fixer still gets the full traceback. Which runner change removes the false green?⌄
MCQ-12An enterprise search service runs hybrid retrieval over 8M AWS documentation chunks: BM25 and HNSW kNN in parallel, RRF fusion, then a cross-encoder rerank of the top 50. p99 is 120 ms against a 150 ms budget (BM25 30, kNN 45, fusion 5, rerank 40). Edits arrive at about 5 chunks/s and must be searchable within 60 s, but a full rebuild takes 4 hours, so the index is rebuilt nightly. Which ingestion design holds both bounds?⌄
MCQ-13A 70B BF16 model (140 GB of weights) is served on two 8×H100 nodes: 3.35 TB/s HBM per GPU, 900 GB/s NVLink inside a node, 50 GB/s InfiniBand between them. It is sharded TP=16 across both nodes; TTFT p99 is 110 ms against a 200 ms budget, but decode runs 22 ms per token (45 tok/s per user against a 100 tok/s target), and profiling puts 12 ms of that in the 160 cross-node all-reduces. Per-user rate is measured with one request in flight. Which layout should replace it?⌄
MCQ-14A 7B BF16 model (14 GB of weights, 1 GB of KV cache at 8K context) serves 8K-token prompts and 128-token answers at batch 1 on one H100 SXM (3.35 TB/s HBM). Prefill p95 is 380 ms and decode 14 ms per token; attention is 22% of prefill and 9% of decode. Enabling FlashAttention-2, which tiles attention so scores stay in SRAM, takes prefill to 300 ms and leaves decode at 14 ms. Which account of those two numbers is right?⌄
MCQ-15A Bedrock agent runs returns chat: sessions reach 60 turns at ~350 tokens each, and the wrapper sends only the last 20 turns because an 8K input cap holds p95 first token at 1.8 s. The customer names `ord_12345678` somewhere in turns 1–25; at turn 34 the refund action is rejected for a missing `order_id`. 31% of these customers have two or more open orders. Which change makes the turn-34 refund send the right `order_id`?⌄
MCQ-16A hybrid RAG service must clear a launch gate of mean recall@10 ≥ 0.90 over a frozen 400-query eval. 120 queries carry an exact identifier token such as `ERR-4021`; the other 280 are natural-language paraphrase. Measured recall@10: BM25 0.94 on identifiers and 0.48 on paraphrase, dense kNN 0.41 and 0.89, RRF (k=60) over both top-10 lists 0.79 and 0.87. Which routing policy clears the gate?⌄
MCQ-17A returns planner calls a 7B model on one 24 GB A10G at 2 req/s. Every prompt is a 2,000-token system prompt and tool-schema block plus about 400 tokens of ticket text, and replies run about 150 tokens. Profiling puts p99 TTFT at 1.4 s (1.15 s prefill, 0.25 s queueing) against a 500 ms budget, per-token time at 28 ms against 50 ms, and KV cache occupancy under 40% of the card. Which change brings TTFT p99 under 500 ms?⌄
MCQ-18A returns assistant handles 18,000 tickets a day. Over 90 days, 96.8% of tickets followed one of four tool sequences; a separate classifier diverts the rest to human review. In the current ReAct loop, 38 of 41 Sev-2 incidents were `issue_refund` firing before `verify_order` returned, all on retries after a payments 504, because the loop re-plans from the last user message and the earlier tool result is no longer in context. Which design should replace it?⌄
MCQ-19A returns agent told 9 of 210 customers over three hours that an out-of-window item was refundable; `check_policy` returned `not_eligible` in all 9, and the reply is free text that nothing compares to the verdict. The previous prompt has the same free-text path, the 400-ticket eval passed at 0.97, three refunds have posted, and the six open quotes are being corrected. Which change stops new wrong answers today?⌄
MCQ-20A returns agent shipped planner-first: one LLM call emitted `plan.steps[]`, which six teams' executors consumed and the pipeline persisted per ticket as the audit record the eval harness replays. Reversal to a step-wise loop took 14 weeks: 1 day to flip the existing flag, 3 days to change the schema, 3 weeks to review codemodded edits at 41 call sites, and 10 weeks re-deriving 4,200 goldens that assert the persisted plan verbatim. Which change to the rollout would have cut that reversal most?⌄
MCQ-21A DSP answers 2M bid requests/s across 200 bidder hosts inside an 80 ms p99 budget after 20 ms of network. Its 7B planner sets per-campaign bid multipliers at 700 ms p99, and the fleet sustains 2,000 planner calls/s; 40,000 campaigns bid in any given minute. Redis reads cost 2 ms p99 at 99.99% availability, the exchange's pacing API is read-only and reports spend on a 60 s lag, and a 3x spike on any one campaign must be capped within 1 s. Which design meets both bounds?⌄
MCQ-22A returns agent drafts a resolution; a reviewer accepts or edits it, and the nightly fine-tune labels an unedited accept as positive. Three weeks in, acceptance rose 71%→92% while 7-day repeat contacts rose 6%→19%, and `fetch_policy` calls fell 0.94→0.21 per ticket as median review time dropped 41 s→12 s. Traces hold every tool call and repeat contacts join back per ticket. Which change to the loop removes the shortcut?⌄
MCQ-23Security telemetry from 300 tenants averages 1 TB/day into a Kinesis stream whose shards each accept 1 MB/s of writes; one tenant produces 40% of the volume, and a detection needs 5 minutes of that tenant's events correlated together. The agent scores one tenant-window per call at 900 ms. Which partitioning and windowing keeps the hot tenant under the shard write limit and still hands the agent complete windows?⌄
MCQ-24A 3-token causal attention layer in float32 computes raw scores `[[2,4,6],[-2,-1,1000],[1000,900,800]]`, multiplies them elementwise by a 0/1 lower-triangular mask, then applies softmax with no max subtraction. What does the layer return for rows 1 and 2?⌄