Part 3 Full Quiz

B is correct. This is a multi-step autonomous task requiring dynamic tool calls based on real-time data — the definition of an agentic scenario. The Azure AI Agent Service lets you register custom function tools (inventory check, carrier API, booking, email) and the agent autonomously decides when and how to call each one, handles results, and loops until complete. A single Chat Completions call cannot call external APIs. Standard Prompt Flow is a fixed linear sequence that cannot adapt dynamically. Logic Apps is a no-code workflow tool without AI reasoning capabilities.

B is correct. Runs in the Azure AI Agent Service are asynchronous. When you call create_run(), the run enters a "queued" state and the API returns immediately — the agent has not yet processed anything. You must poll get_run() in a loop until the status becomes "completed" (or "failed"). Only then can you read the response messages. This is why the while loop polling pattern exists in agent code.

B is correct. Without the @kernel_function decorator, the method is just a regular Python method — Semantic Kernel does not expose it as a tool the LLM can invoke. The decorator is what registers the function and its description with the kernel. Without a description, even if the decorator is present, the LLM won't know when to call it. Plugin name and function name don't need to match. GPT-4o fully supports Semantic Kernel plugins. The Kernel does not need to be re-created.

B is correct. Issuing refunds is a financial, irreversible action — exactly the scenario where Human-in-the-Loop is required. The agent should research and recommend, but a human (customer service agent) must approve each refund before it executes. Temperature 0 makes the model consistent, but a consistently wrong decision is still wrong. Context window size doesn't fix misinterpretation. Semantic Kernel vs Agent Service is an infrastructure choice — neither inherently makes better business decisions.

C is correct. The coordinator agent (unnamed in the scenario) receives tasks, decomposes them, assigns to specialists (MarketAgent, QuantAgent, WriterAgent), collects their results, and synthesises a final output. This is exactly the Orchestrator → Specialists pattern. Sequential Pipeline would have agents running in a fixed order without a coordinator. Group Chat would have all agents debating the same topic together. The specialists here receive independent assignments from the orchestrator, not sequential hand-offs.

B is correct. This is Indirect Prompt Injection — an attacker embeds malicious instructions inside legitimate-looking content (the email) that the agent processes. The agent, unable to distinguish legitimate instructions from hidden attacker instructions in the content, follows them. The mitigation is Prompt Shields with Document Attack detection — this scans all external content for embedded instructions before the agent processes it. This is the #1 security risk for agentic systems and a key exam topic.

B is correct. AutoGen provides max_round (on GroupChat) and max_consecutive_auto_reply (on UserProxyAgent) to set hard limits on how many conversation rounds or auto-replies can occur. When the limit is hit, the conversation terminates rather than continuing indefinitely. Temperature 0 makes outputs deterministic but doesn't stop looping. ALWAYS mode (C) would require human approval at every step — disruptive and defeats the purpose of automation. Switching to Sequential Pipeline (D) changes the whole architecture unnecessarily.

B is correct. The scenario requires cross-session memory — remembering customers across different conversations, not just within one thread. This is long-term memory, implemented with a vector database (Azure AI Search). The agent stores customer preferences and history as embeddings; when a customer returns, the agent queries the vector store to retrieve their relevant history and personalise responses. A system prompt (A) is static. max_tokens (C) controls response length. Creating new threads (D) loses all history since threads are per-session.

B is correct. Semantic Kernel is the right choice when you need custom planning logic with complex conditional decision paths. It provides lower-level SDK access — you can implement custom planners, intercept tool results, apply business rules, and control exactly how the agent reasons and acts. The Azure AI Agent Service is managed — great for standard enterprise agents but the agentic loop is Azure-controlled and less customisable. The services are not equivalent for complex planning. AutoGen is for multi-agent orchestration, not custom planning logic for a single agent.

A is correct. (1) No internet access → private endpoints / VNet integration isolates the agent from the public internet. (2) Traceable decisions → audit logging via Application Insights tracing records every tool call and reasoning step. (3) Human approval before sending → Human-in-the-Loop pauses the agent before delivering trade recommendations. (4) Web agent cannot access order management → least privilege — only give each agent the specific tools it needs for its role, not access to other systems. Option B incorrectly uses temperature=0 for consistency (not an approval control) and content filters (not internet isolation). Options C and D don't map correctly.