CertifyDeckCargando contenido…
AI-500 · Sample questions
10 real questions from the CertifyDeck AI-500 bank, spread across the domains of the official Microsoft Certified: Multi-Agent AI Solutions Expert skills outline, with the correct answer and a full explanation of why the other options are wrong. They are original questions written from the public syllabus, not exam dumps, and they are the same style you will find in the practice tests and mock exams.
Correct answer: A. Add a SummarizationStrategy ahead of the sliding window in the pipeline
A sliding window drops the oldest turns outright, so early facts vanish. SummarizationStrategy uses an LLM to replace older spans with a summary that by default preserves key facts, decisions and user preferences, and placing it before the sliding window in a composed pipeline keeps that context while tokens stay bounded. TruncationStrategy also removes the oldest groups, and the tool-focused strategies only touch tool-call groups, not the user's earlier messages.
Correct answer: C. Tool Input Accuracy
Tool Input Accuracy validates tool call parameters against strict criteria: groundedness, type compliance, format compliance, required parameters, no unexpected parameters and value appropriateness. Tool Call Success only checks whether the calls ran without technical errors, Tool Selection checks that the right tools were chosen, and Tool Output Utilization checks how the agent used the results.
Correct answer: B. Prompts that contain a code matching the pattern are blocked
Setting isRegex to true makes the item a regular expression instead of an exact match, and attaching the list under promptBlocklists with blocking set to true applies it to user prompts and blocks matches. Completions would be filtered only if the list were attached under completionBlocklists. A literal match would require isRegex set to false, and blocking set to true means the request is rejected rather than only annotated.
Correct answer: D. Simulated throttling and timeouts to verify retry and backoff
The Well-Architected testing guidance says to simulate throttling such as HTTP 429 errors, backend timeouts and service unavailability, and to validate that clients handle retries, backoff and circuit breaking; it also recommends simulating tool or agent failures to test reliability. Groundedness measures answer quality, raising quota avoids the failure instead of testing it, and Crescendo is an adversarial safety strategy.
Correct answer: A. Unattended app-only access with the client credentials flow
Agent identities support an unattended, application-only scenario in which the agent acts under its own authority through the OAuth 2.0 client credentials flow, with access governed by its own RBAC role assignments and no human user involved. The on-behalf-of flow and OAuth identity passthrough both require a signed-in user, and placing a key in instructions embeds a secret, which agent identities are designed to avoid.
Correct answer: D. Evaluates the agent's responses as they complete, with a cap on runs per hour
An EvaluationRule with a ContinuousEvaluationRuleAction and the RESPONSE_COMPLETED event type is the continuous evaluation setup: it fires on the filtered agent's response completions, and max_hourly_runs caps the evaluation runs. A daily run over a dataset is the scheduled evaluation, which uses a Schedule with a RecurrenceTrigger instead, and evaluation rules score responses after the fact rather than blocking them like a guardrail.
Correct answer: B. Call await client.configure_azure_monitor() on the chat client
configure_azure_monitor on FoundryChatClient sets up Azure Monitor with the connection string taken from the Foundry project, so traces reach the connected Application Insights resource without wiring exporters yourself. ENABLE_SENSITIVE_DATA only adds prompt and response content to telemetry that is already being exported, and the console or Aspire options send traces to local tools instead of Application Insights.
Correct answer: C. It can reduce prompt overhead, lowering token costs and latency
Fine-tuning adjusts the model weights instead of adding examples to each prompt, so fewer prompt examples are needed, which can lower token costs and latency. The fine-tuning overview states that fine-tuning does not replace retrieval for current information or application-level safety controls, and that benefits are not guaranteed, so a held-out evaluation against the baseline is still needed.
Correct answer: B. Checkpointing, then answer the pending request that is re-emitted on restore
Checkpoints capture the workflow state at superstep boundaries, and pending requests are saved as part of that state; when you restore from a checkpoint, pending requests are re-emitted as request events so you can answer them. Restarting from the initial input replays prior work, which the requirement rules out, and keeping state only in memory doesn't survive a restart. Autonomous mode removes the human gate instead of preserving it.
Correct answer: B. Send an mcp_approval_response with the approval_request_id and the reviewer's decision
With require_approval set to always, the agent emits an mcp_approval_request instead of running the tool; the app reviews the server, tool name and arguments and sends an mcp_approval_response that references approval_request_id with approve set to true or false, linked to the previous response. The tool hasn't run yet, so treating the request as informational skips the gate. Turning approval off or calling the server directly bypasses the reviewer, which fails the requirement.
Answer each question before opening the explanation, and write down why you discarded the other options. On AI-500, most mistakes come from choosing an option that is technically possible but does not meet every requirement in the scenario.
These 10 questions are a taste of the format, not a diagnostic. To know which domains you fail, practice by domain for free and take a timed mock exam that mirrors the official weights and duration.
Start for free, no sign-up or card. Every question includes an explanation like the ones above, and the mock exam reproduces the official AI-500 format.