Skip to main content

SARAVANA KUMAR · TECHNICAL KNOWLEDGE · experienced

Agentic AI: experienced interview questions

20 concise answers with illustrative scenarios and follow-up questions. Use them to explain reasoning, not memorize scripts.

1. How do you prevent agent loops?

Use bounded steps, explicit progress checks and a terminal escalation path, then test repeated tool failures.

Illustrative scenario

An agent keeps searching the same question.

Explain your reasoning

Apply the answer to this scenario: How will a stalled state be detected? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How will a stalled state be detected?

Related worked example and code · Official Agentic AI documentation

2. How do you make tool execution reliable?

Validate schema and semantics, check permissions, set timeouts and classify retryable errors.

Illustrative scenario

A proposed order update has an invalid status.

Explain your reasoning

Apply the answer to this scenario: Should the model repair or the application reject it? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Should the model repair or the application reject it?

Related worked example and code · Official Agentic AI documentation

3. How do you secure approval workflows?

Store approval server-side, bind it to immutable parameters and user identity, enforce expiry and prevent replay.

Illustrative scenario

A recipient changes after payment approval.

Explain your reasoning

Apply the answer to this scenario: Why is matching only an action ID insufficient? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Why is matching only an action ID insufficient?

Related worked example and code · Official Agentic AI documentation

4. How do you evaluate RAG?

Separate retrieval relevance from answer faithfulness and task usefulness, using representative labelled questions.

Illustrative scenario

The answer is wrong despite relevant source material.

Explain your reasoning

Apply the answer to this scenario: Is the failure retrieval, context selection or generation? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Is the failure retrieval, context selection or generation?

Related worked example and code · Official Agentic AI documentation

5. How do you handle tool timeouts?

Treat the external effect as potentially unknown, reconcile by a stable operation key and retry only safely.

Illustrative scenario

A message-sending API times out after accepting a request.

Explain your reasoning

Apply the answer to this scenario: How do you prevent a duplicate message? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How do you prevent a duplicate message?

Related worked example and code · Official Agentic AI documentation

6. How do you design tenant isolation?

Enforce identity and access rules in storage and retrieval layers, not only in prompts.

Illustrative scenario

Two companies use the same assistant platform.

Explain your reasoning

Apply the answer to this scenario: Can metadata supplied by the model bypass the filter? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Can metadata supplied by the model bypass the filter?

Related worked example and code · Official Agentic AI documentation

7. How do you use checkpoints?

Persist resumable workflow state with appropriate identity and versioning, and design side effects for replay.

Illustrative scenario

A workflow resumes after approval.

Explain your reasoning

Apply the answer to this scenario: Which code executes again after resumption? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Which code executes again after resumption?

Related worked example and code · Official Agentic AI documentation

8. When use multi-agent designs?

Use them when a measured decomposition benefit outweighs extra latency, coordination and failure complexity.

Illustrative scenario

Independent review may improve a difficult analysis.

Explain your reasoning

Apply the answer to this scenario: How do you compare it with a single-agent baseline? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How do you compare it with a single-agent baseline?

Related worked example and code · Official Agentic AI documentation

9. How do you prevent secret leakage?

Minimize model-visible secrets, scope tools and redact outputs and logs.

Illustrative scenario

An error message contains a credential.

Explain your reasoning

Apply the answer to this scenario: Which layer removes it before the model sees it? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Which layer removes it before the model sees it?

Related worked example and code · Official Agentic AI documentation

10. How do you choose a model?

Evaluate task success, latency, cost, context needs and failure modes on representative data.

Illustrative scenario

A classifier may not need the most expensive model.

Explain your reasoning

Apply the answer to this scenario: What is the fallback policy? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

What is the fallback policy?

Related worked example and code · Official Agentic AI documentation

11. How do you handle model upgrades?

Version configuration and run regression evaluations before controlled rollout.

Illustrative scenario

A new model changes tool-selection behaviour.

Explain your reasoning

Apply the answer to this scenario: What triggers rollback? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

What triggers rollback?

Related worked example and code · Official Agentic AI documentation

12. How do you control token cost?

Measure useful task outcomes per cost, reduce irrelevant context and avoid unnecessary iterative calls.

Illustrative scenario

An agent repeatedly rereads the same document.

Explain your reasoning

Apply the answer to this scenario: Could caching introduce stale or cross-user data? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Could caching introduce stale or cross-user data?

Related worked example and code · Official Agentic AI documentation

13. How do you validate citations?

Check that each source exists, is authorized and actually supports the associated claim.

Illustrative scenario

An answer cites a real page for an unrelated statement.

Explain your reasoning

Apply the answer to this scenario: How would unsupported claims be flagged? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How would unsupported claims be flagged?

Related worked example and code · Official Agentic AI documentation

14. What belongs in durable memory?

Only justified, authorized information with retention and deletion rules, distinct from transient working state.

Illustrative scenario

A personal assistant remembers a user preference.

Explain your reasoning

Apply the answer to this scenario: How are stale or sensitive memories corrected? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How are stale or sensitive memories corrected?

Related worked example and code · Official Agentic AI documentation

15. How do you design fallbacks?

Provide an honest bounded response or escalation when tools, evidence or confidence are insufficient.

Illustrative scenario

The source system is unavailable.

Explain your reasoning

Apply the answer to this scenario: Why is fabricated completion worse than explicit failure? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Why is fabricated completion worse than explicit failure?

Related worked example and code · Official Agentic AI documentation

16. How do you monitor quality in production?

Combine safe traces, sampled review, feedback and outcome metrics, respecting privacy and sampling bias.

Illustrative scenario

Users stop trusting an otherwise fast assistant.

Explain your reasoning

Apply the answer to this scenario: Which metric reveals unsupported answers? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Which metric reveals unsupported answers?

Related worked example and code · Official Agentic AI documentation

17. How do you test adversarial inputs?

Use authorized test cases covering malicious retrieved text, malformed parameters and access-boundary attempts.

Illustrative scenario

A document instructs a tool to send data elsewhere.

Explain your reasoning

Apply the answer to this scenario: What does a passing test require beyond a refusal sentence? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

What does a passing test require beyond a refusal sentence?

Related worked example and code · Official Agentic AI documentation

18. How do you deploy an agent safely?

Separate configuration, identity, state and tool execution, then roll out with observability and recovery controls.

Illustrative scenario

A staging assistant becomes a production service.

Explain your reasoning

Apply the answer to this scenario: Which credentials and rate limits change? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Which credentials and rate limits change?

Related worked example and code · Official Agentic AI documentation

19. How do you handle evaluation disagreement?

Define the task rubric, inspect disputed cases and combine automated checks with qualified review.

Illustrative scenario

Two graders disagree on answer usefulness.

Explain your reasoning

Apply the answer to this scenario: Can the rubric distinguish style from factual correctness? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

Can the rubric distinguish style from factual correctness?

Related worked example and code · Official Agentic AI documentation

20. What makes a good agent architecture interview answer?

State the goal, trust boundaries, state transitions, failure paths and measurable evaluation before naming frameworks.

Illustrative scenario

Design a document-support assistant.

Explain your reasoning

Apply the answer to this scenario: How would you prove the design works under failure? State your assumptions, the expected outcome and how you would check it before changing a real system.

Common follow-up

How would you prove the design works under failure?

Related worked example and code · Official Agentic AI documentation