AI-500 Questions and Answers

Log the start/completion of each specialist run and inspect the final output with Agent run middleware; record each customer-refund-tool call, including function name and arguments, with Function calling middleware.

The two logging requirements occur at different lifecycle boundaries. Agent run middleware surrounds an agent invocation, so it is the appropriate place to capture run start, run completion, exceptions, and the final agent result. Function calling middleware surrounds tool/function execution and can inspect the selected function, arguments, result, and failures. That makes it the correct boundary for an audit record of each customer-refund-tool invocation. Foundry tracing can provide broader distributed observability, but the question asks for reusable Microsoft Agent Framework middleware at the exact execution points. A knowledge store or memory store does not provide execution interception. Separating the two concerns also supports a cleaner audit model: agent-level middleware records specialist behavior, while function-level middleware records tool use. This aligns with Microsoft ' s middleware model, where cross-cutting concerns such as logging, validation, and exception handling can be implemented once and applied consistently rather than duplicated inside every agent or tool.

Official Microsoft reference: Microsoft Agent Framework - Defining middleware

============================================================

The malicious text is contained in an uploaded email rather than being typed directly as the user ' s prompt. Microsoft classifies malicious instructions embedded in external or retrieved content as an indirect prompt-injection attack. Prompt Shields for documents is designed to detect those attacks in document content before the content can steer the model away from its system instructions. Prompt Shields for user prompts addresses direct attacks originating in the user ' s prompt and therefore targets the wrong attack surface here. Groundedness evaluates whether a response is supported by context; it does not prevent an injected instruction from changing agent behavior. Task Adherence can assess whether an agent follows its task constraints, but it is not the primary control for document-borne prompt injection. Because the source of the attack is the uploaded email, the document-oriented Prompt Shields control is the technically aligned guardrail. The same configuration should be paired with auditable identity, trace, and evaluation data so reviewers can prove which principal acted, which policy was applied, and why a request was allowed or blocked. That is particularly important for production multi-agent systems with external tools.

Official Microsoft reference: Microsoft Foundry guardrails - intervention points and indirect attacks

============================================================

The business rule requires human review for refunds above the stated threshold before payment can continue. A workflow therefore needs a user or operator interaction point that pauses progress and obtains an explicit response. In the workflow options provided, Ask a question is the node that introduces that human interaction. Invoke agent simply delegates work to another agent; it does not itself guarantee human approval. Deliver a message sends information without collecting a decision, and Go to changes control flow without obtaining authorization. Microsoft Agent Framework and Foundry workflow guidance treat human-in-the-loop approval as a first-class orchestration requirement for consequential actions. The decisive requirement is not just to notify a reviewer, but to stop automated progression until the reviewer supplies a decision. Therefore the workflow must include an interactive question/approval step before the refund-processing action can proceed for amounts over the threshold. The implementation should also preserve clear inputs and outputs around this step so that later agents receive only the information they require. This improves debuggability and keeps token, permission, and state growth under control as the workflow becomes more complex.

Official Microsoft reference: Microsoft Agent Framework - Human-in-the-loop workflows

============================================================

Prior claims: memory-store-490; Contact preferences: memory-store-490.

Both requirements describe information that should persist across conversations for the same customer: previous claim information and stable contact preferences. Microsoft Foundry Memory is intended for durable, cross-session facts and summaries that agents can retrieve later, while a knowledge base is for shared reference content rather than user-specific history. The refund-processing and customer-refund tools are operational interfaces, not persistence layers. Using the existing memory store for both categories therefore matches the scenario ' s design objective. In production, the memory search tool should also be scoped to the end user so one customer ' s memory cannot be retrieved for another customer. Microsoft documents scope-based isolation for memory stores and supports a user-derived scope such as `{{$userId}}`. The key distinction is that the same memory store can hold multiple kinds of durable customer memory as long as retrieval is correctly scoped; there is no need to create separate tools or knowledge indexes for these two user-specific memory categories.

Official Microsoft reference: Create and use memory in Foundry Agent Service

============================================================

Fraud-check, Policy-eligibility, and Document-summary are independent specialist activities, so running them concurrently reduces wall-clock time without changing their outputs. The Decision agent is different because the case study states that it depends on the results of those three agents. It therefore must execute only after the independent work completes. Microsoft Agent Framework ' s concurrent orchestration pattern is designed for independent participants whose results can be fanned in later, while dependency-driven work should remain ordered. Running all four agents concurrently would violate the Decision agent ' s data dependency. Running all four sequentially would preserve correctness but add unnecessary latency. Disabling memory would also change functionality unrelated to the performance bottleneck. The optimal graph is therefore a parallel fan-out of the three independent specialists followed by a fan-in and Decision step, which preserves behavior while minimizing end-to-end claim-approval duration. In production, add telemetry and regression tests around this behavior so changes to prompts, models, tools, or orchestration do not silently alter the intended contract. The selected approach is the one that best matches the platform ' s native execution semantics.

Official Microsoft reference: Microsoft Agent Framework - Concurrent orchestration

============================================================

Provision Test, Acceptance, and Production by using a Bicep file applied through the CI/CD pipeline in Azure Pipelines; block promotion when quality falls below the threshold by using Foundry Agent Evaluator as a pipeline release gate.

The release lifecycle must satisfy two independent requirements: repeatable infrastructure deployment across DTAP environments and an automated quality gate before promotion. Bicep is Azure ' s infrastructure-as-code language, and Microsoft documents deploying Bicep through Azure Pipelines so the same declarative resources can be promoted consistently across subscriptions. For the promotion decision, an automated Foundry evaluation is the appropriate control because response quality is an AI-system property that should be measured against a defined threshold before the next stage is released. A manual portal deployment would weaken repeatability, while a deployment-only gate would not measure model or agent behavior. The important design point is that infrastructure provisioning and AI quality validation are separate controls in the same pipeline: Bicep establishes environment parity, and the evaluation gate prevents a technically deployable but behaviorally degraded agent version from advancing. From a security and governance perspective, the control should be enforced at the narrowest platform boundary that can deterministically block or constrain the action. Relying only on prompt text is weaker because the model can still be induced to behave unexpectedly.

Official Microsoft reference: Deploy Bicep files with Azure Pipelines; AI-500 Study Guide

============================================================