# Agentic AI Interview – Experienced Hire - Part 2 ## INTERVIEW CONTEXT ### Target Roles - Backend Engineers - ML Engineers - Platform Engineers - AI Engineers (3+ years) ### Interview Focus - System design thinking - Production experience - Failure handling and trade-offs ## Q3: Memory Design Decisions ### Interview Question - How do you design memory for an agentic system? ### What the Interviewer Is Testing - System design thinking - Privacy and compliance awareness - Failure mode handling ### Strong Answer Signals - Separates short term and long term memory - Explains when NOT to store memory - Mentions audit and replay needs ### Weak Answer Signals - Just naming a vector database - No privacy or retention discussion ### Follow Up Questions - What if the memory store is down? - How do you handle sensitive data? ## Q4: Failure Handling and Retries ### Interview Question - An external API call fails. What happens next? ### What the Interviewer Is Testing - Production readiness - Idempotency awareness - Operational experience ### Strong Answer Signals - Partial failure handling - Safe retries via idempotency - Timeouts and backoff strategy - Observability and tracing ### Weak Answer Signals - Retry three times with try catch - No observability discussion ### Follow Up Questions - How do you prevent cascading failures? - How would you debug this at 3 AM? ## Q5: Evaluation and Human in the Loop ### Interview Question - How do you know your agent is working correctly? ### What the Interviewer Is Testing - Evaluation as a continuous process - Risk based automation decisions ### Strong Answer Signals - Automated evaluation plus human approval gates - Mentions regression and monitoring - Explains which actions require humans ### Weak Answer Signals - Test with a few examples - Human reviews everything ### Follow Up Questions - How do you detect drift? - How do you measure quality over time? ## INTERVIEW SUCCESS PATTERNS ### Ask clarifying questions first ### Explain trade-offs explicitly ### Think out loud ### Discuss failure modes proactively