Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 21

Question 200

In an architecture review, the key requirement is to reduce latency when separate read-only lookups have no ordering dependency. Which solution directly satisfies it?

  1. Orchestrate a deterministic sequence around model steps
  2. Handle streamed generation with cancellation
  3. Run independent tool requests concurrently where safe
  4. Route conditional branches using rules

Explanation

Correct Answer

C:

For the stated requirement to reduce latency when separate read-only lookups have no ordering dependency, parallelizing independent operations shortens critical-path latency but is unsafe when one action requires another's output; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to execute a known approval workflow where inputs must be validated before a model drafts a response, explicit workflow stages make ordering and failure handling reproducible; free-form agent planning is unnecessary for a fixed process; this does not directly meet the required outcome to reduce latency when separate read-only lookups have no ordering dependency.

B:

For a system whose objective is to show partial assistant output to a user while stopping work after the request is canceled, streaming delivers incremental output; cancellation and backpressure prevent orphaned inference work and stale interface content; this does not directly meet the required outcome to reduce latency when separate read-only lookups have no ordering dependency.

D:

For a system whose objective is to choose among distinct business flows based on a validated account status, rules engines offer predictable decisions for simple structured conditions, while language models handle unstructured interpretation; this does not directly meet the required outcome to reduce latency when separate read-only lookups have no ordering dependency.

Question 201

A product team must remember a customer's explicit notification preference across separate sessions. What is the most appropriate Azure AI implementation?

  1. Publish a versioned agent endpoint for stable clients
  2. Retrieve consented durable preferences when required
  3. Summarize old dialogue for context-window management
  4. Expire temporary secrets and tool results

Explanation

Correct Answer

B:

For the stated requirement to remember a customer's explicit notification preference across separate sessions, persistent preference storage is distinct from a single chat transcript and requires consent, access control, and deletion workflows; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to let several applications consume an agent while its implementation changes under controlled release, versioned published endpoints decouple client contracts from internals and support staged promotion when the platform supports it; this does not directly meet the required outcome to remember a customer's explicit notification preference across separate sessions.

C:

For a system whose objective is to retain key decisions after a long conversation becomes too large to replay verbatim, a validated rolling summary compresses history while preserving commitments and identifiers; careless compression can discard safety constraints; this does not directly meet the required outcome to remember a customer's explicit notification preference across separate sessions.

D:

For a system whose objective is to avoid retaining authentication tokens or temporary PII inside agent memory longer than needed, short-lived state and purpose-based retention protect secrets and sensitive data; summarization is not a substitute for removal; this does not directly meet the required outcome to remember a customer's explicit notification preference across separate sessions.

Question 202

What test should engineers run after they add code execution for bounded data calculations, rather than relying solely on a successful API response?

  1. Inject mid-workflow failures and audit the final reconciled business state
  2. Trace agent runs for actual knowledge-tool invocation and check support for each source-grounded answer
  3. Compare computed totals against independently verified reference calculations
  4. Measure field-level extraction accuracy on labeled scans and low-quality images

Explanation

Correct Answer

C:

The accepted change is to add code execution for bounded data calculations; this evaluation checks whether it can let an agent aggregate a supplied numeric CSV where exact computation is important, since a sandboxed execution tool performs calculations deterministically; executable code needs strict isolation and resource limits.

Incorrect Answers

A:

This test concerns the ability to recover from a partially completed workflow when a later operation fails, for which a compensating action undoes or offsets previously committed steps where a transaction cannot be atomic across services; it does not establish whether the change to add code execution for bounded data calculations works for the current release.

B:

This test concerns the ability to invoke governed document search as a tool at the point the agent needs evidence for an answer, for which an agent run needs explicit retrieval calls and returned context before composing a grounded response; model parameters alone do not fetch current documentation; it does not establish whether the change to add code execution for bounded data calculations works for the current release.

D:

This test concerns the ability to enable an agent to read key fields from photographed or scanned paperwork, for which document extraction tools convert visual text and layout into structured data suitable for subsequent agent reasoning; it does not establish whether the change to add code execution for bounded data calculations works for the current release.

Question 203

A new production AI solution must require one specialist's validated output before another starts its dependent work. Which approach should the team select first?

  1. Use a dispatcher agent to select a specialist
  2. Limit information passed between specialist agents
  3. Run task-completion evaluations with ground truth
  4. Use a sequential handoff between specialist agents

Explanation

Correct Answer

D:

For the stated requirement to require one specialist's validated output before another starts its dependent work, a sequential handoff makes intermediate contracts explicit; parallel execution would violate a data dependency; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to route finance, operations, and security questions to different bounded expert agents, a router delegates to an appropriate specialist with a narrow tool set; a single all-powerful agent is harder to govern; this does not directly meet the required outcome to require one specialist's validated output before another starts its dependent work.

B:

For a system whose objective is to stop one agent from receiving private data that another agent needed only for its subtask, scoped messages and redacted handoff contracts minimize sensitive context propagation across multiple agents; this does not directly meet the required outcome to require one specialist's validated output before another starts its dependent work.

C:

For a system whose objective is to determine whether the agent actually achieved a multi-step business goal, task-success tests compare final system state with expected outcomes; prose fluency does not prove work was completed; this does not directly meet the required outcome to require one specialist's validated output before another starts its dependent work.

Question 204

A project needs to complete all of these distinct tasks: (1) let several applications consume an agent while its implementation changes under controlled release; and (2) avoid giving every agent separate uncontrolled credentials for the same external service. Select one suitable capability for each task. Choose TWO.

  1. Apply idempotency to write-capable tools
  2. Separate agent identity from the human requester
  3. Decommission obsolete agent versions safely
  4. Scope shared tool access centrally
  5. Publish a versioned agent endpoint for stable clients
  6. Use a content-extraction tool for uploaded forms

Explanation

Correct Answers

D:

For the stated requirement to avoid giving every agent separate uncontrolled credentials for the same external service, centralized governed tool access can standardize authentication and logging while each agent remains least privileged; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to avoid giving every agent separate uncontrolled credentials for the same external service, while the other deliverables require separate capabilities.

E:

For the stated requirement to let several applications consume an agent while its implementation changes under controlled release, versioned published endpoints decouple client contracts from internals and support staged promotion when the platform supports it; this is the mechanism that the system must implement. This choice satisfies the specific deliverable to let several applications consume an agent while its implementation changes under controlled release, while the other deliverables require separate capabilities.

Incorrect Answers

A:

For a system whose objective is to prevent double charging when an agent retries after a network timeout, an idempotency key ties a logical transaction to one effect even if the same request is retried; this does not directly meet the required outcome to avoid giving every agent separate uncontrolled credentials for the same external service.

B:

For a system whose objective is to enforce what both the service and the delegated user are authorized to do, a workload identity authenticates the agent service, while user-aware authorization limits operations on behalf of a caller; this does not directly meet the required outcome to avoid giving every agent separate uncontrolled credentials for the same external service.

C:

For a system whose objective is to remove old versions that retain deprecated tools without disrupting active clients, a controlled retirement checks clients, revokes unused credentials, and archives relevant audit records before deletion; this does not directly meet the required outcome to avoid giving every agent separate uncontrolled credentials for the same external service.

F:

For a system whose objective is to enable an agent to read key fields from photographed or scanned paperwork, document extraction tools convert visual text and layout into structured data suitable for subsequent agent reasoning; this does not directly meet the required outcome to avoid giving every agent separate uncontrolled credentials for the same external service.

Question 205

After implementing a workflow to retrieve consented durable preferences when required, the team needs a robust release gate. Which check should it choose?

  1. Verify recall across sessions alongside correct deletion and user isolation
  2. Test pronoun resolution across turns and verify session boundaries
  3. Audit per-agent tool entitlements and cross-service invocation traces
  4. Check recall of critical decisions and track context-token savings over long sessions

Explanation

Correct Answer

A:

The accepted change is to retrieve consented durable preferences when required; this evaluation checks whether it can remember a customer's explicit notification preference across separate sessions, since persistent preference storage is distinct from a single chat transcript and requires consent, access control, and deletion workflows.

Incorrect Answers

B:

This test concerns the ability to resolve 'those two records' by retaining the preceding selection context, for which conversation tracking gives referential continuity across turns, but the context should be bounded to the correct user session; it does not establish whether the change to retrieve consented durable preferences when required works for the current release.

C:

This test concerns the ability to avoid giving every agent separate uncontrolled credentials for the same external service, for which centralized governed tool access can standardize authentication and logging while each agent remains least privileged; it does not establish whether the change to retrieve consented durable preferences when required works for the current release.

D:

This test concerns the ability to retain key decisions after a long conversation becomes too large to replay verbatim, for which a validated rolling summary compresses history while preserving commitments and identifiers; careless compression can discard safety constraints; it does not establish whether the change to retrieve consented durable preferences when required works for the current release.

Question 206

Before approving the proposed action to give agents bounded execution budgets, which measurable acceptance check should be required?

  1. Run no-result test cases and verify termination under configured budgets
  2. Measure tool precision, forbidden-call rate, and unnecessary invocation count
  3. Force timeouts and confirm the payment ledger records one transaction
  4. Inject mid-workflow failures and audit the final reconciled business state

Explanation

Correct Answer

A:

The accepted change is to give agents bounded execution budgets; this evaluation checks whether it can stop self-triggering tool loops from consuming unlimited tokens or external calls, since step, time, and spending limits create enforceable failure boundaries regardless of the prompt plan.

Incorrect Answers

B:

This test concerns the ability to detect unnecessary or dangerous tool calls even when its final prose looks plausible, for which tool-selection evaluation examines intermediate actions and authorization fit, not only the final natural-language answer; it does not establish whether the change to give agents bounded execution budgets works for the current release.

C:

This test concerns the ability to prevent double charging when an agent retries after a network timeout, for which an idempotency key ties a logical transaction to one effect even if the same request is retried; it does not establish whether the change to give agents bounded execution budgets works for the current release.

D:

This test concerns the ability to recover from a partially completed workflow when a later operation fails, for which a compensating action undoes or offsets previously committed steps where a transaction cannot be atomic across services; it does not establish whether the change to give agents bounded execution budgets works for the current release.

Question 207

An Azure AI workload passes basic health probes but has this issue: An extended planning agent exceeds the model's context budget when every old message is replayed. Which fix targets it?

  1. Summarize older turns, keep current facts verbatim, and validate the retained state
  2. Load only authorized durable preferences into the current session context
  3. Invoke an authorized OCR-and-layout extraction tool before applying business rules
  4. Exclude transient secrets from memory and enforce state expiration policies

Explanation

Correct Answer

A:

The fault signal is that an extended planning agent exceeds the model's context budget when every old message is replayed; the targeted fix works because a validated rolling summary compresses history while preserving commitments and identifiers; careless compression can discard safety constraints, restoring the ability to retain key decisions after a long conversation becomes too large to replay verbatim.

Incorrect Answers

B:

If the failing function were to remember a customer's explicit notification preference across separate sessions, this change could be relevant because persistent preference storage is distinct from a single chat transcript and requires consent, access control, and deletion workflows; the actual symptom is that an extended planning agent exceeds the model's context budget when every old message is replayed, which calls for a different repair.

C:

If the failing function were to enable an agent to read key fields from photographed or scanned paperwork, this change could be relevant because document extraction tools convert visual text and layout into structured data suitable for subsequent agent reasoning; the actual symptom is that an extended planning agent exceeds the model's context budget when every old message is replayed, which calls for a different repair.

D:

If the failing function were to avoid retaining authentication tokens or temporary PII inside agent memory longer than needed, this change could be relevant because short-lived state and purpose-based retention protect secrets and sensitive data; summarization is not a substitute for removal; the actual symptom is that an extended planning agent exceeds the model's context budget when every old message is replayed, which calls for a different repair.

Question 208

Stakeholders want a Foundry-based application able to find where a lengthy agent plan diverges from intended workflow. Which choice implements that requirement?

  1. Trace step-level agent decisions and failures
  2. Run task-completion evaluations with ground truth
  3. Evaluate whether an agent chooses appropriate tools
  4. Maintain task state apart from transient model output

Explanation

Correct Answer

A:

For the stated requirement to find where a lengthy agent plan diverges from intended workflow, per-step tracing captures model requests, tool parameters, statuses and transitions; a final response log hides the path taken; this is the mechanism that the system must implement.

Incorrect Answers

B:

For a system whose objective is to determine whether the agent actually achieved a multi-step business goal, task-success tests compare final system state with expected outcomes; prose fluency does not prove work was completed; this does not directly meet the required outcome to find where a lengthy agent plan diverges from intended workflow.

C:

For a system whose objective is to detect unnecessary or dangerous tool calls even when its final prose looks plausible, tool-selection evaluation examines intermediate actions and authorization fit, not only the final natural-language answer; this does not directly meet the required outcome to find where a lengthy agent plan diverges from intended workflow.

D:

For a system whose objective is to resume a long-running workflow safely after an inference retry, durable workflow state records completed steps and pending approvals; relying only on the model's last text risks duplicate side effects; this does not directly meet the required outcome to find where a lengthy agent plan diverges from intended workflow.

Question 209

In an architecture review, the key requirement is to detect unnecessary or dangerous tool calls even when its final prose looks plausible. Which solution directly satisfies it?

  1. Use a supervisor to coordinate parallel worker agents
  2. Give agents bounded execution budgets
  3. Run task-completion evaluations with ground truth
  4. Analyze agent errors by failure class
  5. Evaluate whether an agent chooses appropriate tools

Explanation

Correct Answer

E:

For the stated requirement to detect unnecessary or dangerous tool calls even when its final prose looks plausible, tool-selection evaluation examines intermediate actions and authorization fit, not only the final natural-language answer; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to combine independent research subtasks into a coherent final response with oversight, a supervisor can assign independent tasks, collect outputs, and apply cross-checks before presenting an answer; this does not directly meet the required outcome to detect unnecessary or dangerous tool calls even when its final prose looks plausible.

B:

For a system whose objective is to stop self-triggering tool loops from consuming unlimited tokens or external calls, step, time, and spending limits create enforceable failure boundaries regardless of the prompt plan; this does not directly meet the required outcome to detect unnecessary or dangerous tool calls even when its final prose looks plausible.

C:

For a system whose objective is to determine whether the agent actually achieved a multi-step business goal, task-success tests compare final system state with expected outcomes; prose fluency does not prove work was completed; this does not directly meet the required outcome to detect unnecessary or dangerous tool calls even when its final prose looks plausible.

D:

For a system whose objective is to prioritize corrections across retrieval misses, malformed tools, policy blocks, and model planning errors, failure classification links observed symptoms to the component that can be improved; a single aggregate error rate obscures causes; this does not directly meet the required outcome to detect unnecessary or dangerous tool calls even when its final prose looks plausible.

Question 210

An engineer receives this incident report: A frontend must be edited every time the developer changes the agent's model or instructions. What should be changed?

  1. Migrate consumers, disable the stale agent endpoint, and revoke its credentials
  2. Route approved computation to an isolated code tool with validated file inputs
  3. Publish a stable agent interface and manage revision selection in deployment configuration
  4. Use an approved shared tool integration with centralized access policy and secret handling

Explanation

Correct Answer

C:

The fault signal is that a frontend must be edited every time the developer changes the agent's model or instructions; the targeted fix works because versioned published endpoints decouple client contracts from internals and support staged promotion when the platform supports it, restoring the ability to let several applications consume an agent while its implementation changes under controlled release.

Incorrect Answers

A:

If the failing function were to remove old versions that retain deprecated tools without disrupting active clients, this change could be relevant because a controlled retirement checks clients, revokes unused credentials, and archives relevant audit records before deletion; the actual symptom is that a frontend must be edited every time the developer changes the agent's model or instructions, which calls for a different repair.

B:

If the failing function were to let an agent aggregate a supplied numeric CSV where exact computation is important, this change could be relevant because a sandboxed execution tool performs calculations deterministically; executable code needs strict isolation and resource limits; the actual symptom is that a frontend must be edited every time the developer changes the agent's model or instructions, which calls for a different repair.

D:

If the failing function were to avoid giving every agent separate uncontrolled credentials for the same external service, this change could be relevant because centralized governed tool access can standardize authentication and logging while each agent remains least privileged; the actual symptom is that a frontend must be edited every time the developer changes the agent's model or instructions, which calls for a different repair.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests