5.3. Practice Questions
Q1. Your agent must answer questions from a partner-facing product manual stored in a VNet-isolated Azure AI Search service. Citations must link to source documents. What must be true of the index? (Choose two.)
A. It uses integrated vectorization B. It contains a URL or metadata_storage_path field C. Semantic ranking is disabled D. It combines all manuals into a single searchable index per connection requirement
Answer: A and B. The knowledge source requires a vector index built with integrated vectorization (A), and citations require a URL-bearing field (B). C is backwards — semantic ranking is supported if pre-configured. D misreads the rule: one index per connection, not all content in one index.
Q2. A logistics company's Copilot Studio agent must delegate carrier-capacity checks to a partner's agent running on a non-Microsoft platform. What should you use?
A. Child agent B. Connected Copilot Studio agent C. A2A protocol integration D. MCP server connection
Answer: C. Cross-platform, cross-organization agent collaboration is A2A's exact job, discovered via the partner's agent card. A and B require the Copilot Studio ecosystem; D connects an agent to tools, not to a peer agent.
Q3. Users report the agent invokes the "create ticket" flow without ever asking for the ticket category, and the confirmation number never appears in the conversation. Which two parts of the flow's configuration should you examine?
Answer: the flow's input parameters (category isn't declared/described as a required input, so the orchestrator never collects it) and its output parameters (the confirmation number isn't declared as an output, or the topic isn't consuming it). This is the parameter contract from 3.2.3 — both directions are broken.
Q4. An insurance agent's claims topic must answer only from the claims index using the firm's approved Azure OpenAI deployment, while the rest of the agent answers from SharePoint knowledge. What should you configure?
A. Add the claims index as agent-level knowledge B. Configure the claims topic's generative answers node data source with the Azure OpenAI connection C. Replace the agent's model in settings D. Add a keyword trigger to block non-claims sources
Answer: B. Node-level data sources take priority over agent-level knowledge and scope grounding to the topic; the Azure OpenAI connection satisfies the approved-deployment requirement. A can't restrict which topic uses it and doesn't touch the model; C isn't how per-topic generation is controlled; D isn't a real mechanism.
Q5. During evaluation, factual answers keep failing even though reviewers agree they're correct — the agent phrases them differently each run. The team also wants to confirm the answers come from the HR knowledge source. What should you change? (Choose two.)
A. Switch the text checks to Compare Meaning B. Switch to Exact Match with a longer expected answer C. Add a Tool Use method to the affected cases D. Lower the pass rate threshold to 20%
Answer: A and C. The failures are a method-variability mismatch — Compare Meaning scores semantic equivalence across phrasings (A). Verifying the retrieval route is Tool Use's job (C). B tightens the mismatch; D games the score instead of measuring correctly.
Q6. A scheduled agent flow stopped running mid-month. Investigation shows no errors in the flow logic. Per Copilot Studio's billing model, what is the most likely cause?
Answer: exhausted prepaid Copilot Studio capacity — flow runs consume capacity per action, and when capacity is consumed, new runs are blocked until it's available. The fix is capacity planning/monitoring, not flow debugging.
Q7. You must deploy an agent solution nightly from dev to test with no one signed in, and require the compliance officer's approval before any production deployment. Which two features do you implement?
Answer: a service principal identity for the unattended nightly pipeline deployment, and a pipeline extension adding an approval gate on the production stage. Manual import answers are the distractor pattern; the approval belongs in the pipeline, not around it.
Q8. An agent for anonymous website visitors must answer questions about published product specs. A stakeholder asks to also let it show each visitor "their order history." What do you tell them?
Answer: order history requires knowing who the user is — impossible under no-auth. Either the audience stays anonymous and the agent serves public content only, or an OAuth-based authentication is configured so users sign in and per-user data can be honored. The two requirements can't coexist in one unauthenticated experience.
