McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 30

Question 290

An agent refreshes short-lived OAuth access tokens for external APIs. Developers propose storing them in long-term user memory so calls are easier next month. Which policy is correct?

  1. Keep transient API authorization tokens out of persistent agent memory.
  2. Copy every token into user-visible conversation history.
  3. Embed tokens in permanent preference vectors so the model can retrieve them.
  4. Disable token expiry so long-term memory can continue using one key.

Explanation

Correct Answer

A:

Access tokens have expiry and secrecy requirements incompatible with generic semantic memory stores.

Incorrect Answers

B:

Chat histories are not an appropriate secret store.

C:

Memory retrieval may expose secrets and retain expired credentials.

D:

Prolonging credential lifetime expands compromise risk and ignores secure OAuth practices.

Question 291

An agent repeatedly invokes optional tools during one user request until its token quota is exhausted. What per-turn guardrail is needed?

  1. Increase the maximum allowed calls without any termination policy.
  2. Apply capped per-turn token and tool-call budgets.
  3. Remove traces so repeated calls are not visible.
  4. Use longer prompts to compensate for tool exhaustion.

Explanation

Correct Answer

B:

Budgets keep autonomous exploration within predictable execution and cost limits.

Incorrect Answers

A:

Unbounded operation can amplify cost and latency.

C:

Concealment does not limit actual execution.

D:

More prompt tokens can worsen the same budget problem.

Question 292

An agent deleted a record after a lookup; the incident team needs to establish which tool step preceded the delete. Which telemetry is essential?

  1. Record only the final natural-language message that the record was deleted.
  2. Trace the order of tool calls to diagnose incorrect agent actions.
  3. Ask the model to reconstruct the execution from memory after the incident.
  4. Count daily model tokens but discard operation identities.

Explanation

Correct Answer

B:

Execution traces reconstruct decision order and dependencies for the problematic action.

Incorrect Answers

A:

The answer does not prove which tools were invoked or in what sequence.

C:

Generated recollection is not authoritative activity evidence.

D:

Aggregate usage cannot identify a specific destructive step.

Question 293

A customer-support agent appears friendly but fails to complete multi-step tasks. Which evaluation design captures the missing quality?

  1. Score output readability and formatting while omitting task completion and tool-result correctness.
  2. Test agent task completion, tool outcomes and required constraints.
  3. Measure answer politeness and completion time, without checking whether external actions succeeded.
  4. Count how many agent tools exist in the configuration.

Explanation

Correct Answer

B:

Agent success requires real task execution and compliance, not only pleasant response style.

Incorrect Answers

A:

Fluent agent responses can conceal unsuccessful tool execution, so operational outcome checks matter.

C:

Friendly language and low latency cannot prove that the agent actually completed the requested workflow.

D:

Tool inventory size says nothing about whether tasks are completed correctly.

Question 294

Two agent instruction versions use the same tools. Team A claims version B is better at selecting the correct business operation. What should be compared?

  1. Judge whichever version emits more tool calls as better.
  2. Let both versions handle different untracked customer populations without comparable evaluation.
  3. Compare only the number of tokens in each version's instructions.
  4. Evaluate tool choice against labeled expected tool calls.

Explanation

Correct Answer

D:

Labeled selection cases distinguish correct routing from merely fluent final answers.

Incorrect Answers

A:

Extra calls can indicate unnecessary or unsafe actions.

B:

Noncomparable samples cannot isolate routing quality differences.

C:

Instruction length is not a direct quality metric for tool selection.

Question 295

An agent executes tools based on responses from a remote web page. A test page inserts hostile instructions to send payroll data externally. Which QA is needed?

  1. Give the test agent live unrestricted payroll export rights.
  2. Treat fixture instructions as system-level because they came from a tool.
  3. Test an agent against fabricated hostile tool outputs.
  4. Test only ordinary friendly pages with no injection content.

Explanation

Correct Answer

C:

Indirect-injection evaluations test whether lower-trust results can redirect protected tool behavior.

Incorrect Answers

A:

Adversarial evaluation should avoid needless real-world sensitive side effects.

B:

Trusted tool transport does not elevate untrusted returned text.

D:

Benign tests do not exercise hostile tool responses.

Question 296

A support agent reports that it updated CRM records, but tool traces show no CRM invocation at all. What failure has occurred?

  1. A confident generated statement proves that the CRM record changed.
  2. The agent fabricated completion: no CRM tool invocation occurred.
  3. Count the number of action verbs in the answer as a transaction audit.
  4. Create a fake tool result afterward so the logs match the response.

Explanation

Correct Answer

B:

A model can describe an action it did not perform; only provider results establish execution.

Incorrect Answers

A:

Confidence wording cannot replace operational evidence.

C:

Language markers do not show real tool side effects.

D:

Invented evidence conceals the underlying failure.

Question 297

An agent has tried a database lookup three times with temporary failures. Further attempts would exceed operational limits. What should it do?

  1. Stop retries at budget exhaustion and report the unresolved action.
  2. Invent the missing database result to finish the answer.
  3. Continue retrying indefinitely until the user closes the session.
  4. Broaden the agent's permissions automatically because the service failed.

Explanation

Correct Answer

A:

Bounded recovery prevents loops and informs the user when an external dependency remains unavailable.

Incorrect Answers

B:

Fabricated data is not a successful lookup.

C:

Unlimited retries waste capacity and can increase dependency stress.

D:

Temporary availability errors do not authorize privilege escalation.

Question 298

A document search tool returns no items accessible to the current employee. The agent is asked for the current policy. What controlled failure response fits?

  1. Return a high-similarity passage without applying the caller’s permitted document scope.
  2. Generate a plausible policy number and cite a nonexistent page.
  3. Report no accessible records instead of inventing an answer.
  4. Use an unrestricted administrator query to obtain any policy record.

Explanation

Correct Answer

C:

Absence-aware retrieval respects both evidentiary and permission boundaries.

Incorrect Answers

A:

Relevance ranking cannot override tenant or role-based access restrictions.

B:

Fictional references make an unsupported answer look authoritative.

D:

The employee's missing access is not permission to bypass security trimming.

Question 299

A multi-agent team combines legal and financial specialist findings. Reviewers need to trace every conclusion to its underlying documents. What must the orchestrator retain?

  1. Let specialist agents generate citation references from their summaries instead of passing through retrieved source IDs.
  2. Record source provenance in multi-agent synthesized reports.
  3. Delete all source references once a specialist produces a confident sentence.
  4. Use the number of specialists as the citation list.

Explanation

Correct Answer

B:

Preserved provenance prevents evidence loss when multiple generated summaries are merged.

Incorrect Answers

A:

A generated locator is unverifiable unless it resolves to an actual retrieved record.

C:

Confidence is not a replacement for auditable evidence.

D:

Agent count does not identify the documents supporting claims.

Question 300

Two agents can access sensitive customer records while collaborating on one support case. Auditors need evidence of cross-agent reads and writes. What should be logged?

  1. Log only the final answer length after the support case ends.
  2. Record principal/agent identity, authorized record access, timestamp, purpose and sanitized outcomes.
  3. Give agents permission to delete their own access traces.
  4. Store full raw customer secrets in public debug traces for transparency.

Explanation

Correct Answer

B:

A protected action log makes sensitive access traceable across agent boundaries.

Incorrect Answers

A:

Response length cannot show which sensitive record was accessed.

C:

Self-modifiable records undermine audit integrity.

D:

Excess data exposure is not needed for accountability.

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