Agentic AI interview questions
AI engineer roles increasingly ask about agents: systems where a language model decides which steps to take, calls tools, reads the results and keeps going until the task is done. Interviewers want to know whether you can build one that is reliable, safe and affordable, not just a demo.
The questions below go from core concepts to production concerns, with what a strong answer should include.
What interviewers assess
- What an agent is, and when a simple workflow is the better choice
- Tool calling, planning, memory and how the agent loop works
- Multi-agent designs and how agents hand work to each other
- Evaluation, guardrails, cost and latency in production
Core concepts questions
What is an AI agent, and how is it different from a chatbot or a fixed LLM workflow?
What a good answer covers: An agent uses a model in a loop to choose actions (tool calls), observe the results and decide the next step until a goal is met. A workflow follows steps the developer fixed in code. Strong answers say workflows are often better when the steps are known, since they are cheaper and more predictable.
Explain how tool calling (function calling) works.
What a good answer covers: You describe tools with a name, description and input schema; the model returns a structured request to call one; your code runs it and sends back the result; the model continues. Mention validating inputs and that the model never runs code itself.
What is the ReAct pattern?
What a good answer covers: The model alternates between reasoning and acting (tool calls), using each observation to choose the next step. Compare it with planning everything up front then executing.
How do agents use memory?
What a good answer covers: Short-term memory is the conversation and tool results in the context window; long-term memory is stored outside (a database or vector store) and retrieved when relevant. Discuss summarising or trimming context as it grows.
What makes a good tool description for an agent?
What a good answer covers: A clear name, when to use it and when not to, precise parameter descriptions, examples, and useful error messages. Tool design often matters more than the prompt.
Design questions
When would you use several agents instead of one?
What a good answer covers: When tasks need clearly different expertise, tools or context, or can run in parallel. Costs: more tokens, harder debugging, and errors passed between agents. One well-equipped agent is often enough.
Describe an orchestrator-worker (supervisor) design.
What a good answer covers: A lead agent breaks the task down, delegates sub-tasks to specialised workers, and combines their results. Discuss how much context each worker gets and how failures come back to the lead.
How do you stop an agent from looping forever or wandering off task?
What a good answer covers: Step and token budgets, timeouts, clear stopping criteria, detecting repeated actions, and returning a partial result with an explanation.
How do you decide which actions need a human to approve them?
What a good answer covers: Based on risk and reversibility: reading is usually safe; sending emails, payments, deleting data or anything public needs confirmation or strict limits.
How would you give an agent access to company systems such as a CRM or database?
What a good answer covers: Through narrow tools with least-privilege credentials, scoped per user, with input validation and audit logs. Protocols such as MCP standardise how tools are exposed to agents.
Production and evaluation questions
How do you evaluate an agent?
What a good answer covers: A set of realistic tasks with checkable outcomes, measuring task success, number of steps, cost and latency; inspecting full traces; and running it again after every prompt, tool or model change to catch regressions.
What is prompt injection, and why is it more dangerous for agents?
What a good answer covers: Instructions hidden in content the agent reads (a web page, email or document) try to take control. With tools, that can lead to real actions such as leaking data. Defences: treat retrieved content as data, limit tool permissions, require confirmation for risky actions.
An agent works in testing but fails often in production. How do you investigate?
What a good answer covers: Look at traces of failed runs, group failures by cause (wrong tool, bad input, missing context, model error), fix the biggest group, add those cases to the evaluation set.
How do you control the cost and latency of an agent?
What a good answer covers: Smaller models for simple steps, prompt caching, fewer and better tools, parallel tool calls, trimming context, and caching tool results.
What would you log and monitor for an agent in production?
What a good answer covers: Every step and tool call with inputs and outputs (minus sensitive data), tokens, cost, latency, errors, task outcome and user feedback.
Common mistakes
- Using an agent where a simple, predictable workflow would do.
- Describing a framework's features instead of how the agent loop actually works.
- No answer for how you would measure whether the agent is any good.
- Ignoring security: giving an agent broad permissions and trusting everything it reads.
How to prepare
- Build a small agent yourself with two or three tools, without a framework first, so you understand the loop.
- Then try one framework (for example LangGraph, CrewAI or a model provider's agent SDK) and compare.
- Write ten test tasks for your agent and measure how many it completes.
- Prepare to talk through one failure you saw and how you fixed it.
Reading questions only goes so far: the real test is answering them aloud, under follow-up questions. Career Scan builds a mock interview from your own resume and the job, then shows exactly which answers need work.
Frequently asked questions
Do I need to know a specific agent framework?
It helps to have used one, but interviewers care more that you understand tool calling, the agent loop, evaluation and safety. Those carry across frameworks.
Are agentic AI questions asked in ML engineer interviews too?
Increasingly, especially for AI engineer and applied ML roles that build products on large language models.
What projects help most for agentic AI roles?
A working agent that does something useful with real tools, plus evidence that you tested it: an evaluation set, failure analysis and cost numbers.