Agentic AI interview questions
“Agentic” is doing a lot of work in job descriptions right now, and it means something narrower in an interview than it does in marketing: a system where the model chooses what to do next and then acts, in a loop, with tools that touch the real world. Everything hard about it follows from those two words, chooses and acts.
62 questions here touch agents, and they are deliberately not filed together: they sit across 5 different topics, because an agent question is sometimes a security question and sometimes a judgement question wearing a systems costume. Below they are grouped by what the interviewer is actually testing.
Grounded in real GenAI, LLM, and AI/ML engineering interview loops and written to a senior-engineer editorial bar.
Should this be an agent at all?
The first and most-failed question. Agents cost latency, money and determinism, and most problems presented as agent problems are a retrieval problem or a workflow with three fixed steps. Interviewers want to hear you decline the agent when a single call would do, then justify it when it would not.
- When do you build an agent instead of a single LLM call, and how do you keep a multi-step agent reliable?RAG & Agent System Design · hard
- A complex agent scores 15% better on your benchmark than a simple RAG pipeline. Which do you ship?Behavioral & Project Deep-Dives · medium
- Given a new LLM use case, how do you decide between prompting, RAG, fine-tuning, and tools/agents?LLM & GenAI Fundamentals · medium
- What is the Plan-and-Execute agent pattern, and how does it compare to ReAct?RAG & Agent System Design · medium
- When should an agent write code instead of emitting JSON tool calls?RAG & Agent System Design · medium
Design one, end to end, on a whiteboard
The system-design round, and the one people underprepare because they have only ever built agents in a notebook. The interviewer is listening for the boring parts: where state lives, what happens on a restart, how a human gets back in the loop, and what it costs per task at a thousand users.
- Design an autonomous coding agent that resolves GitHub issues end to end (plan, edit, test, iterate).System Design for AI in Production · hard
- Design a deep research agent that answers complex questions by searching and synthesizing many sources.System Design for AI in Production · hard
- Design an agentic workflow platform where users compose LLM agents that call tools and run for minutes.System Design for AI in Production · expert
- Design a computer-use agent that operates a browser to complete tasks (book travel, fill forms) reliably.System Design for AI in Production · hard
- Design a multi-agent customer support system with escalation to humans.System Design for AI in Production · hard
Will it actually finish?
Every demo agent works. The interview is about the long tail: how it knows it is done, what stops it burning a budget, what happens when it dies forty minutes into an hour-long task, and where the context goes when the conversation outgrows the window.
- How does an agent know when it is done?RAG & Agent System Design · medium
- How do you enforce a hard per-task budget so an agent stops before it overruns?RAG & Agent System Design · medium
- How do you design state and checkpointing for an agent that runs for an hour and might crash?RAG & Agent System Design · hard
- How do you manage memory and context for a long-running conversational agent?RAG & Agent System Design · hard
- Your agent takes 90 seconds to complete a task. How do you make it fast?RAG & Agent System Design · medium
Is it safe to let it act?
The round that has grown fastest. The moment an agent reads untrusted text and can also call a tool, prompt injection stops being a curiosity and becomes an exfiltration path. Expect to be pushed on identity and least privilege, not just on filtering.
- A tool-using agent reads untrusted web content. How do you defend against prompt injection?AI Security, Privacy & Governance · hard
- What is indirect prompt injection, and why is it so dangerous for RAG and agents?AI Security, Privacy & Governance · hard
- Your agent calls tools on behalf of users. How do you design its identity, credentials, and authorization?AI Security, Privacy & Governance · hard
- How do you let an AI agent execute code safely (sandboxing)?RAG & Agent System Design · hard
- How do tool-result and memory poisoning attacks compromise an AI agent, and how do you defend?AI Security, Privacy & Governance · hard
Can you prove it works?
Agents fail in trajectories, not in single outputs, so the usual metrics do not transfer. This is the round strong builders most often lose, because they can describe what they shipped but not how they knew it was good enough to ship.
- How would you evaluate a coding agent, and why is a single SWE-bench score not enough to trust it?LLM & GenAI Fundamentals · hard
- What is agent reflection / self-correction, and does it actually improve agent performance?RAG & Agent System Design · medium
- When do you use a multi-agent system, and what orchestration patterns and pitfalls matter?RAG & Agent System Design · hard
- Your team merges 50 agent-written PRs a week and review is the bottleneck. Design the review workflow.RAG & Agent System Design · expert
If the vocabulary is the problem, start here instead
Questions are for rehearsing out loud. If you would rather learn the mechanism first, these are the concept pages the questions above assume. The first one is free.
Where this sits in the rest of the bank
The full topic, in the order to work through it, is RAG & agent system design. The security half lives in AI security, privacy & governance. If you do not know which to open first, take the diagnostic and let the ones you fumble decide for you.