Part 9c — Question Bank: Domain 3, Integration (Q46–Q74)
Original practice questions; not official or leaked. Objectives quoted from Exam Guide v1.0 §6 D3. Domain 3 is the heaviest at 19%.
Q46 · D3 · Objective: Evaluate tool/agent configuration for capability bloat · Difficulty: Intermediate
Scenario: An internal assistant exposes 45 tools to every conversation. Traces show frequent wrong-tool choices and 7,000 tokens of tool definitions per request.
Question: Which change is best?
A. Increase max_tokens.
B. Scope tools to the task or role and add on-demand tool discovery for the rest; re-run tool-selection evals.
C. Add a prompt line: "Choose tools carefully."
D. Switch to a model with a larger context.
Correct answer: B
Explanation: Fewer, scoped tools reduce confusion and tokens; discovery loads detail only when needed.
Why not the others: A and D don't reduce confusion or cost; C is not an enforcement of anything.
Architectural principle: Least capability plus progressive discovery.
Ref: Prep Lesson 2; Agent Skills (progressive disclosure)
Q47 · D3 · Objective: Evaluate tool/agent configuration for capability bloat · Difficulty: Intermediate
Scenario: A business-intelligence assistant only answers read-only questions, but its database tool runs under a role that can also INSERT and DELETE "in case we need it later." Question: What is the best correction? A. Log all write statements for later audit. B. Add a system-prompt rule: "Never write to the database." C. Ask the model to confirm before writing. D. Run the tool under a read-only role (and restrict tables/rows), creating a separate, gated path if writes are ever needed. Correct answer: D Explanation: Removing unneeded privilege eliminates the risk; monitoring and prompts only compensate. Why not the others: A is detective; B and C depend on model compliance. Architectural principle: Preventive controls (least privilege) beat detective or compensating ones. Ref: Prep Lessons 2–3
Q48 · D3 · Objective: Evaluate tool/agent configuration for capability bloat · Difficulty: Advanced
Scenario: The tool list includes search_docs, find_documents, and lookup_docs with nearly identical descriptions; the agent uses them inconsistently and results vary.
Question: What should you do?
A. Consolidate into one well-specified tool (or clearly differentiate the purposes) with precise, non-overlapping descriptions and schemas; then re-evaluate.
B. Add more tools with other names.
C. Randomize tool order each request.
D. Add examples showing all three used at once.
Correct answer: A
Explanation: Overlap causes ambiguous selection; consolidation or crisp differentiation improves reliability.
Why not the others: B worsens it; C adds noise; D increases confusion.
Architectural principle: Tool design: distinct purposes, clear contracts.
Ref: Tool use docs
Q49 · D3 · Objective: Evaluate tool/agent configuration for capability bloat · Difficulty: Intermediate
Question (Select TWO): Which observations most strongly indicate capability bloat in an agent? A. Production traces show many configured tools are never called by real workflows. B. A high prompt-cache hit rate. C. Frequent wrong-tool selections among tools with similar names and descriptions. D. A short system prompt. E. Low p95 latency. Correct answers: A, C Explanation: Unused tools and confusable tools are classic symptoms of over-tooling. Why not the others: B, D, and E are neutral or positive signs. Architectural principle: Use traces to find and remove unneeded capability. Ref: Exam Guide D3
Q50 · D3 · Objective: Analyze authentication and authorization requirements to identify security gaps · Difficulty: Advanced
Scenario: A CRM agent uses a single administrator token for all lookups. A junior rep asks about a restricted VIP account and the agent returns it. Question: Which is the most important fix? A. Add "respect permissions" to the prompt. B. Hide the account name in the response. C. Propagate the end user's identity (on-behalf-of / user-scoped tokens) so the CRM enforces that user's entitlements. D. Reduce the number of CRM fields the model can see. Correct answer: C Explanation: The shared super-user token makes the agent a confused deputy; backend enforcement per user closes the gap. Why not the others: A relies on the model; B and D are cosmetic or partial. Architectural principle: Authorize with the user's identity at the system of record. Ref: MCP authorization spec; Prep Lesson 3
Q51 · D3 · Objective: Analyze authentication and authorization requirements to identify security gaps · Difficulty: Advanced
Scenario: A get_invoices tool takes tenant_id and customer_id as arguments supplied by the model.
Question: What is the primary risk and fix?
A. Latency; cache invoice results.
B. The model (or an injected instruction) can request another tenant's data; derive tenant from the authenticated session and validate all arguments server-side.
C. Token cost; shorten argument names.
D. Formatting; require uppercase IDs.
Correct answer: B
Explanation: Model-supplied identifiers are untrusted input; authorization must use authenticated context.
Why not the others: The other options don't address cross-tenant access.
Architectural principle: Never trust model-supplied authority; validate server-side.
Ref: Prep Lesson 3
Q52 · D3 · Objective: Analyze authentication and authorization requirements to identify security gaps · Difficulty: Expert
Scenario: An HTTP-based MCP server accepts any valid OAuth token from the company's identity provider, including tokens issued for other internal APIs, and forwards them downstream. Question: What is the gap and the best fix? A. No gap; any company token is fine. B. Switch the server to stdio so OAuth isn't needed. C. Hide the server URL. D. Tokens must be issued for, and validated against, the MCP server as the intended audience; the server must not pass through tokens meant for other resources, and each tool must still apply its own data authorization. Correct answer: D Explanation: Audience binding prevents token confusion; passing tokens through breaks trust boundaries; protocol-level auth doesn't replace per-tool authorization. Why not the others: A ignores confused-deputy risk; B changes transport and credential handling, not the gap; C is obscurity. Architectural principle: Audience-bound tokens, no passthrough, and per-resource authorization. Ref: https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization
Q53 · D3 · Objective: Analyze authentication and authorization requirements to identify security gaps · Difficulty: Advanced
Scenario: A shared vector index serves several tenants. The app retrieves top-k chunks, sends them to Claude, then removes sentences from the answer that "look confidential." Question: What is the most important correction? A. Apply tenant/ACL filters at retrieval time so unauthorized chunks never reach the model. B. Increase k to find more safe chunks. C. Add a disclaimer to the answer. D. Fine-tune the model to avoid confidential text. Correct answer: A Explanation: Post-hoc filtering is unreliable and the model has already seen the data. Why not the others: B, C, and D don't enforce access boundaries. Architectural principle: Entitlement filtering before context. Ref: Prep Lessons 2–3
Q54 · D3 · Objective: Analyze authentication and authorization requirements to identify security gaps · Difficulty: Advanced
Question (Select TWO): Which are authorization gaps? A. One shared API key embedded in the system prompt and used for every user's requests. B. User-scoped, short-lived tokens exchanged per request. C. Entitlement checks performed only after retrieved text has been sent to the model. D. A broker that validates tool arguments against a schema. E. Audit logs recording user and agent identity. Correct answers: A, C Explanation: Shared embedded credentials and after-the-fact filtering both defeat least privilege. Why not the others: B, D, and E are good controls. Architectural principle: Identify where authority is granted and where it's checked. Ref: Exam Guide D3/D5
Q55 · D3 · Objective: Evaluate accuracy-latency trade-offs and justify configuration decisions · Difficulty: Intermediate
Scenario: A consumer chat feature needs the first words to appear within a second; answers are moderate in length; quality on the eval set is acceptable with a mid-tier model.
Question: Which configuration best supports the goal?
A. Use batch processing.
B. Add a verification pass to every answer.
C. Stream responses, cache the static prefix, keep max_tokens appropriate, and use the lowest model tier that clears quality targets.
D. Use the largest model and rely on streaming alone.
Correct answer: C
Explanation: Streaming improves perceived latency; caching and right-sizing reduce time-to-first-token and cost.
Why not the others: A is asynchronous; B adds latency; D increases latency and cost.
Architectural principle: Combine streaming, caching, and right-sizing for interactive latency.
Ref: Streaming docs; prompt caching docs
Q56 · D3 · Objective: Evaluate accuracy-latency trade-offs and justify configuration decisions · Difficulty: Advanced
Scenario: A second "verification" model call doubles latency. Errors in low-risk draft replies cost little; errors in refund approvals are expensive. Question: What is the best configuration decision? A. Remove verification everywhere. B. Apply verification (and stronger models) selectively to high-risk actions, and skip or lighten it for low-risk drafts, justified with eval data. C. Apply verification everywhere regardless of risk. D. Use verification only when the user complains. Correct answer: B Explanation: Risk-tiered configuration spends latency/cost where errors are costly. Why not the others: A and C ignore risk differences; D is reactive. Architectural principle: Allocate accuracy effort by error cost. Ref: Prep Lesson 2
Q57 · D3 · Objective: Evaluate accuracy-latency trade-offs and justify configuration decisions · Difficulty: Advanced
Scenario: Month-end spikes cause 429 errors. Retries create duplicate refund tickets in the downstream system. Question: Which combination is best? A. Increase retries to 20 with no delay. B. Remove retries. C. Ask the model not to refund twice. D. Exponential backoff with jitter honoring retry-after, a queue with concurrency limits, and idempotency keys on write operations. Correct answer: D Explanation: Backoff and queuing manage load; idempotency prevents duplicate side effects on retry. Why not the others: A worsens the storm; B drops legitimate requests; C is not a control. Architectural principle: Resilience with idempotent writes. Ref: https://docs.claude.com/en/api/rate-limits
Q58 · D3 · Objective: Analyze observability challenges and select monitoring strategies at scale · Difficulty: Intermediate
Scenario: Support reports more wrong answers, yet dashboards show 99.9% availability and normal latency. Question: What is the monitoring gap? A. Add quality signals: sampled automated evals on live traffic, user feedback, retrieval and citation metrics, linked to traces. B. Add more uptime checks. C. Add a status page. D. Increase log retention to forever. Correct answer: A Explanation: Availability metrics don't measure answer quality. Why not the others: B and C measure availability; D adds cost without insight. Architectural principle: Monitor quality, not just uptime. Ref: Prep Lesson 2
Q59 · D3 · Objective: Analyze observability challenges and select monitoring strategies at scale · Difficulty: Advanced
Scenario: Engineers need prompts and tool arguments in logs for debugging, but they may contain personal data. Question: Which approach is best? A. Log everything raw with broad access. B. Disable logging. C. Redact or tokenize sensitive fields at the edge, limit retention and access, and keep immutable audit logs separate from debug traces. D. Log only on weekends. Correct answer: C Explanation: Balances debuggability with privacy and audit requirements. Why not the others: A creates a privacy risk; B blinds operations; D is arbitrary. Architectural principle: Observability with privacy by design. Ref: Prep Lessons 2–3
Q60 · D3 · Objective: Analyze observability challenges and select monitoring strategies at scale · Difficulty: Advanced
Scenario: At 50M requests/day, storing full traces for all traffic is cost-prohibitive. Question: Which strategy is best? A. Store nothing. B. Collect metrics for all traffic; sample full traces while always retaining errors, low-confidence outputs, flagged cases, and security events. C. Store only the first 1,000 requests per day. D. Store traces only for the largest customers. Correct answer: B Explanation: Aggregated metrics plus targeted sampling keep visibility into the cases that matter. Why not the others: A, C, and D create blind spots. Architectural principle: Sample intelligently; always keep exceptions. Ref: Prep Lesson 2
Q61 · D3 · Objective: Design a RAG pipeline with appropriate chunking and indexing strategies · Difficulty: Intermediate
Scenario: Policy documents have headings and tables. Fixed 200-token chunks split tables mid-row, and answers cite fragments without context. Question: Which change is best? A. Increase k to 50. B. Use a bigger model. C. Remove the tables. D. Use structure-aware chunking (respect headings/tables) with section context and metadata attached to each chunk. Correct answer: D Explanation: Chunks should be coherent and self-describing for both retrieval and answering. Why not the others: A adds noise; B doesn't fix broken chunks; C loses data. Architectural principle: Chunk by structure with context. Ref: Prep Lesson 2
Q62 · D3 · Objective: Design a RAG pipeline with appropriate chunking and indexing strategies · Difficulty: Advanced
Scenario: After one team re-embedded new documents with a different embedding model, retrieval quality for old documents dropped. Question: What is the likely cause and fix? A. Embeddings from different models are not comparable; re-embed the corpus consistently (or maintain separate indexes per embedding model) and version the embedding config. B. The LLM forgot the old documents. C. The documents are too short. D. The system prompt is too long. Correct answer: A Explanation: Mixed embedding spaces corrupt similarity search. Why not the others: B, C, and D don't explain a drop tied to the embedding change. Architectural principle: Consistent, versioned embedding and index configuration. Ref: Prep Lesson 2
Q63 · D3 · Objective: Design a RAG pipeline with appropriate chunking and indexing strategies · Difficulty: Intermediate
Scenario: The assistant sometimes cites a superseded benefits policy after a new one was published. Question: Which design addresses this? A. Tell users to double-check. B. Version documents with effective dates, remove or tombstone superseded chunks on update, and monitor index freshness. C. Lower retrieval k to 1. D. Add "use the newest policy" to the system prompt. Correct answer: B Explanation: Freshness is an indexing responsibility; stale chunks must be retired. Why not the others: A shifts burden; C risks recall; D can't know which chunk is newest without metadata. Architectural principle: Version, tombstone, and monitor the index. Ref: Prep Lesson 2
Q64 · D3 · Objective: Apply retrieval strategies matched to data shape and query pattern · Difficulty: Intermediate
Question (Select TWO): Users search by exact part numbers like "AX-4410-B" and semantic search returns loosely related items. Which changes help most? A. Combine semantic search with keyword/BM25 (hybrid retrieval). B. Raise generation temperature. C. Use exact-match filters or a structured lookup for identifier fields. D. Shorten the system prompt. E. Switch to a larger generation model. Correct answers: A, C Explanation: Identifier queries need lexical or structured matching, not purely semantic similarity. Why not the others: B, D, and E don't change retrieval behavior. Architectural principle: Match retrieval to query pattern. Ref: Prep Lesson 2
Q65 · D3 · Objective: Apply retrieval strategies matched to data shape and query pattern · Difficulty: Intermediate
Scenario: Analysts ask, "What was Q3 EMEA revenue by product line?" The data lives in a warehouse. Question: Which approach is best? A. Embed the warehouse rows and use vector search. B. Paste the full table into the prompt. C. Use a validated, read-only structured query path (e.g., parameterized SQL or API) with schema allow-lists and row limits, and have Claude explain the result. D. Ask Claude to estimate from memory. Correct answer: C Explanation: Aggregations belong in query engines; constrained access keeps them safe. Why not the others: A and B are inaccurate or wasteful; D is unreliable. Architectural principle: Use structured retrieval for structured aggregates. Ref: Prep Lesson 2
Q66 · D3 · Objective: Apply retrieval strategies matched to data shape and query pattern · Difficulty: Advanced
Scenario: Questions such as "Which suppliers of components used in Product X had incidents last year?" require following relationships across several sources. Question: Which retrieval approach is most appropriate? A. One vector search with k=3. B. Keyword search on the question text only. C. Fine-tune the model on supplier names. D. Iterative (agentic) retrieval or graph-style queries across sources, with budgets and citations. Correct answer: D Explanation: Multi-hop questions need sequential retrieval steps or relationship-aware queries. Why not the others: A and B rarely connect hops; C doesn't solve freshness or traceability. Architectural principle: Multi-hop patterns need iterative or relational retrieval. Ref: Prep Lesson 2; Building effective agents
Q67 · D3 · Objective: Evaluate progressive discovery vs. monolithic context strategy · Difficulty: Intermediate
Scenario: A 40-page, rarely changing product guide is used in almost every support query. Question: What is the simplest effective design? A. Put the guide in a cached static prefix (monolithic context) and verify cache hits. B. Build a complex multi-index RAG system. C. Make the agent search the guide with an iterative tool each time. D. Summarize it into 200 words. Correct answer: A Explanation: Small, stable, always-relevant content fits caching well; RAG adds complexity without benefit. Why not the others: B and C over-engineer; D loses detail. Architectural principle: Don't add retrieval complexity when caching suffices. Ref: Prompt caching docs
Q68 · D3 · Objective: Apply retrieval strategies matched to data shape and query pattern · Difficulty: Advanced
Scenario: End-to-end answer quality on a RAG system is 78%. You suspect retrieval but aren't sure. Question: What should you do first? A. Switch to the largest model. B. Measure retrieval separately (e.g., recall@k on labeled questions) before changing the generator; consider reranking or hybrid search if recall is low. C. Rewrite the system prompt. D. Increase temperature. Correct answer: B Explanation: Component-level evals localize the bottleneck. Why not the others: A, C, and D change the wrong layer without evidence. Architectural principle: Evaluate components separately. Ref: Exam Guide D4
Q69 · D3 · Objective: Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent) · Difficulty: Intermediate
Scenario: Three applications need the same ticketing integration now, and two more teams will add applications next quarter. Question: Which mechanism is best? A. Each app writes its own direct API wrapper. B. Use a CLI on developers' laptops. C. Expose a ticketing agent per app. D. Build one MCP server for ticketing with centralized authentication/authorization and shared tool definitions. Correct answer: D Explanation: Reuse across many clients is MCP's sweet spot. Why not the others: A multiplies maintenance and inconsistent security; B isn't a server-side integration; C is unnecessary. Architectural principle: Shared, reusable integrations → MCP. Ref: https://modelcontextprotocol.io
Q70 · D3 · Objective: Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent) · Difficulty: Intermediate
Scenario: One internal application calls a single stable internal pricing API on the hot path; latency is tight and no other consumers exist. Question: Which is most appropriate? A. Adopt an agent-to-agent protocol. B. Introduce an MCP server and gateway for future flexibility. C. A direct tool function that calls the API with schema validation and user-scoped authorization. D. A shell CLI called from the model. Correct answer: C Explanation: One consumer and tight latency favor the simplest path with the fewest hops. Why not the others: A and B add hops and operations without need; D broadens risk. Architectural principle: Don't add integration layers without a reuse or governance benefit. Ref: Prep Lesson 2
Q71 · D3 · Objective: Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent) · Difficulty: Advanced
Scenario: A partner exposes an autonomous claims-adjudication agent behind its own API. Your agent should delegate a goal ("assess this claim") rather than call fine-grained functions. Question: What is the best integration approach? A. Agent-to-agent delegation, with scoped identity propagation, defined input/output contracts, timeouts, result verification, and audit. B. Scrape the partner's UI. C. Copy the partner's prompts into your system. D. Give the partner agent your admin token. Correct answer: A Explanation: Delegating a goal to another autonomous agent fits agent-to-agent patterns, with controls for trust and verification. Why not the others: B is brittle; C is infeasible and inappropriate; D violates least privilege. Architectural principle: Delegate with contracts, verification, and bounded identity. Ref: Exam Guide D3
Q72 · D3 · Objective: Evaluate connection protocols and select the appropriate integration mechanism (MCP, API/CLI, agent-to-agent) · Difficulty: Advanced
Question (Select TWO): A developer wants to install a popular community MCP server that can read internal files. Which are appropriate steps before adopting it? A. Install the latest version automatically on every start. B. Review the source and tool descriptions, then pin a reviewed version. C. Give it an administrator token to avoid permission errors. D. Run it with least-privilege credentials, restricted file scope, and network egress limits in a sandbox. E. Trust it because it has many stars. Correct answers: B, D Explanation: Third-party servers are supply-chain risk; tool descriptions can carry injection; least privilege contains damage. Why not the others: A removes review; C violates least privilege; E is not a security control. Architectural principle: Vet, pin, and sandbox integrations. Ref: MCP security guidance; Prep Lesson 3
Q73 · D3 · Objective: Evaluate progressive discovery vs. monolithic context strategy · Difficulty: Advanced
Scenario: An enterprise platform has 200 internal tools; any given task needs about five. Loading all definitions bloats prompts and confuses selection. Question: Which design is best? A. Load all 200 definitions into every request and raise the token limit. B. Provide a small core set plus a discovery mechanism (search/skill index) so relevant tools load on demand. C. Randomly sample 20 tools each time. D. Hard-code five tools for all tasks. Correct answer: B Explanation: Progressive discovery scales as catalogs grow. Why not the others: A bloats; C risks missing needed tools; D can't serve varied tasks. Architectural principle: Progressive discovery for large capability catalogs. Ref: Agent Skills; Prep Lesson 2
Q74 · D3 · Objective: Evaluate progressive discovery vs. monolithic context strategy · Difficulty: Intermediate
Scenario: A focused assistant uses six stable tools that are relevant to almost every conversation. Question: Which design is most appropriate? A. Build a tool-search system with dynamic loading. B. Split the six tools across six agents. C. Load all six tool definitions directly (monolithic) and cache them in the stable prefix. D. Remove half the tools randomly. Correct answer: C Explanation: A small, stable, always-relevant set doesn't need discovery overhead. Why not the others: A and B add complexity; D harms capability. Architectural principle: Match complexity to catalog size and variability. Ref: Prep Lesson 2