Agentic AI · Module 37 of 41 · 7 concepts · about 10 min

Interview: Beginner Questions

How agent interviews are structured, how to answer a definition question well, and the questions asked in screening rounds: what agents are, how tools and the loop work, the basics of memory, RAG and prompts, and why limits matter.

At work: The first twenty minutes of almost every interview for a role that touches agents. These answers decide whether the harder questions are ever asked.

How Agentic AI interviews are structured

What it is

Interviews for roles involving agents usually have four kinds of round:

Screening: definitions and basics. 'What is an agent? How does tool calling work?'
Technical depth: how things work and fail. 'Explain ReAct. How would you implement memory? What is agentic RAG?'
Coding: write a small part live. A tool executor, an agent loop, a retry with backoff, a simple evaluation.
System design: 'Design an agent that…', for forty-five to sixty minutes.

Senior rounds add scenario questions ('your agent is doing X, what do you do?') and questions about projects you have built.

Why it matters

Knowing the format tells you what each question is really testing.

Screening tests whether you can explain an idea plainly.
Depth tests whether you know the trade-offs and the failure modes.
Coding tests whether you know what is under the frameworks.
Design and scenarios test judgement: the simplest thing that works, and what you do when it breaks.
At every level, interviewers listen for the same habits: limits in code, success checked in the world, and a path to a human.

In the real world: A hiring manager describes the pattern: 'by minute ten I know whether they have read about agents or built one, because I have asked what stops the loop'.

Visual example

Rounds

The modules listed are the ones to revise for each.

The four interview rounds and what each tests

round	typical question	tests	revise
Screening	What is an AI agent?	Clear, plain explanation	Modules 1 to 3, 37
Technical depth	How does ReAct work, and how does it fail?	Mechanism and trade-offs	Modules 5 to 22, 38
Coding	Write an agent loop with a tool	What is under the frameworks	Modules 9, 10, 34
System design	Design a refund agent	Judgement; safety; evaluation	Modules 23 to 33, 39, 41
Scenario	It keeps calling the same API. Now what?	Diagnosis and repair	Modules 25, 28, 40

Preparing only definitions. Past the first round, almost every question is 'and what goes wrong with that?' or 'how would you know?'.

How to answer a definition question

What it is

A reliable shape for any 'what is X?' question, in about forty seconds:

One sentence that is a complete definition in plain words.
One concrete example.
One contrast or trade-off: what it is not, or what it costs.
Then stop, and let the interviewer choose where to go deeper.

Why it matters

The shape shows three things at once: that you understand the idea, that you have seen it used, and that you know its limits.

Starting with a framework name or a list of buzzwords signals the opposite.
Stopping is a skill. A four-minute answer to a screening question leaves no time for the questions that distinguish you.
The same shape works for 'explain this to a product manager', which is a common variant.

In the real world: Interviewers often ask the same opening question of every candidate; the answers that stand out are shorter than the average, not longer.

Visual example

Answer shape

Definition, example, trade-off, stop.

Forty seconds for a definition

One sentenceA plain definition
One exampleA concrete case
One trade-offWhat it costs or is not
StopLet them choose the follow-up

Q: What is an AI agent? (weak)

“An agent is an autonomous AI system built with frameworks like LangChain that uses LLMs, tools, memory and planning to perform complex tasks through reasoning.” A list of nouns. It does not say what the thing does.

Q: What is an AI agent? (strong)

“A language model that uses tools in a loop: it decides the next step from what it has found so far, my code carries that step out, and it repeats until the task is done or a limit stops it. For example, a support agent that looks up an order, checks the policy and issues the refund, instead of only describing the policy. The cost is that a mistake is now a wrong action, not a wrong sentence, so the limits have to be in code.”

Reciting a memorised paragraph. Interviewers follow up immediately on whichever word you used that you cannot explain. Use only words you can define.

Beginner questions: agents and agentic systems

What it is

The opening questions of a screening round. They test whether you can say what an agent is, and what it is not, without jargon.

Why it matters

Say each answer aloud before reading it.

The first sentence should be a complete answer on its own.
The rest adds one example or one trade-off.
If an answer here surprises you, the module named in the question's topic is where to go back.

In the real world: A screening call covers perhaps six of these in ten minutes; candidates who answer each in under a minute get to the interesting questions.

Q: What is an AI agent?

A language model that uses tools in a loop, choosing its own next step from what it has observed, until the goal is met or a limit is reached. The model decides; your code runs the tools and enforces the limits.

Q: What makes an AI system agentic?

The model, not the programmer, decides what happens next, and that decision can depend on the result of the last step. In practice: it is goal-directed, uses tools, reacts to feedback and persists over several steps. It is a matter of degree, not a yes or no.

Q: Agent versus chatbot?

A chatbot answers in one model call. An agent takes actions over many calls and decides when the task is finished. A chatbot tells you the refund policy; an agent issues the refund.

Q: Agent versus LLM?

An LLM maps text to text and has no memory, no tools and no loop. An agent is an LLM placed in a loop with tools, instructions and state. The LLM is one component of the agent.

Q: Agent versus RAG?

RAG is a fixed pipeline: retrieve once, then answer. An agent decides whether, what and how many times to retrieve, and can also act. RAG is often one of an agent's tools.

Q: Agent versus workflow?

In a workflow the code fixes the steps and the model fills them in. In an agent the model chooses the steps. Workflows are predictable and cheap; agents handle the unplanned. Most real systems are a workflow with an agent step where flexibility is needed.

Q: What is the agent loop?

Observe, reason, decide, act, and observe again. The model reads the history, chooses a tool or a final answer, the code runs the tool, and the result is added to the history for the next turn.

Q: When should you not use an agent?

When the steps are known in advance, when success cannot be checked, or when a mistake is costly and irreversible with no approval step. A single call or a workflow is cheaper, faster and easier to test.

Q: What are levels of autonomy?

How much the agent does without a person: assist, suggest, act with approval, act then report, act alone. The level is set per action by how costly and reversible a mistake is.

Q: Give an example of a real agentic system.

A coding agent: it reads the repository, edits files, runs the tests and fixes what failed, then opens a change for human review. It works because tests give precise feedback and review catches mistakes.

Defining an agent as 'an autonomous AI'. Autonomous how, deciding what, bounded by what? Say the mechanism.

Practise it: Design a Self-Correcting Code Agent, Design an AI Customer Support Agent

Beginner questions: tools and function calling

What it is

The second most common block. It tests whether you know that the model only asks, and your code does.

Why it matters

Say each answer aloud before reading it.

The first sentence should be a complete answer on its own.
The rest adds one example or one trade-off.
If an answer here surprises you, the module named in the question's topic is where to go back.

In the real world: The follow-up to 'how does function calling work?' is nearly always 'so where would you check permissions?'. The answer is: in your code, between the request and the execution.

Q: What are tools?

Functions in your code that the model can ask to have run. Each has a name, a description and an input schema. They let the model read live data, compute exactly and take actions.

Q: Why do agents need tools?

A model alone knows only its training data, cannot calculate reliably and cannot change anything. Tools supply current facts, exact computation and real effects.

Q: How does function calling work?

You send tool definitions with the conversation. The model replies with a structured request: the tool's name and JSON inputs. Your code runs the function and returns the result with a matching id. The model then continues. The model never executes anything itself.

Q: What is a tool schema?

A JSON Schema describing a tool's inputs: names, types, which are required, allowed values. The model reads it to form a call, and your code uses it to reject a bad one.

Q: How does the model choose a tool?

By reading the names and descriptions. So a description should say what the tool does, when to use it and when not to. Most wrong choices are description problems.

Q: What happens when a tool fails?

The error is returned to the model as the tool's result, with a message it can act on. Passing faults such as timeouts are retried by code with backoff. Permission failures are not retried.

Q: What is the difference between a read tool and a write tool?

A read tool looks something up and is safe to repeat. A write tool changes something, so it needs permission checks, limits, and protection against being run twice.

Q: Can the model call several tools at once?

Yes. It can request several in one reply; the code runs independent ones in parallel and returns all the results together. That cuts both waiting time and the number of turns.

Saying 'the model calls the API'. It requests; your code calls.

Beginner questions: memory, RAG and prompts

What it is

Basic questions on what the model knows and how you tell it things.

Why it matters

Say each answer aloud before reading it.

The first sentence should be a complete answer on its own.
The rest adds one example or one trade-off.
If an answer here surprises you, the module named in the question's topic is where to go back.

In the real world: An interviewer who hears 'the model remembers the user' stops and asks how; the candidate who says 'it does not, my code resends or retrieves it' has passed a quiet test.

Q: Does an LLM remember previous conversations?

No. Each call sees only what is sent in that call. Conversation memory is your code resending the history; long-term memory is a store your code reads and writes.

Q: What is a context window?

The maximum amount of text, measured in tokens, a model can read in one call. Everything the model knows about the task must fit in it, and every token is paid for.

Q: Short-term versus long-term memory?

Short-term is the current task's history, resent each turn and gone when the task ends. Long-term is information saved outside the model and retrieved in later sessions.

Q: What is RAG?

Retrieval-augmented generation: search your documents for passages relevant to the question, put them in the prompt, and have the model answer from them. It grounds answers in your data.

Q: What is a system prompt?

The standing instructions sent with every call: who the agent is, how it should work, its limits and what to do when stuck. It is the same for every task; the user's message is the part that changes.

Q: What is a hallucination?

The model stating something false as if it were true. In an agent it can also be a claimed action that never happened, or an invented id. The defences are grounding answers in tool results and verifying them in code.

Q: What is a structured output?

A reply in a fixed machine-readable shape, usually JSON matching a schema, so that code can act on it. Routing, tool inputs, plans and verdicts all need it.

Q: What is a token?

A piece of a word. Models read and are billed in tokens; a page of English is roughly five hundred.

Beginner questions: limits and safety basics

What it is

Even at screening level, interviewers check that you know an agent needs bounds.

Why it matters

Say each answer aloud before reading it.

The first sentence should be a complete answer on its own.
The rest adds one example or one trade-off.
If an answer here surprises you, the module named in the question's topic is where to go back.

In the real world: The simplest safety question, 'how would you limit a refund?', has one answer interviewers are waiting for: 'in the tool, not in the prompt'.

Q: How do you stop an agent running for ever?

With limits set in code: a maximum number of steps, a budget of tokens or money, a time limit and detection of repeated calls. When one triggers, the agent stops and hands over.

Q: What is a guardrail?

A check that stops an agent doing, saying or accepting something it should not: on the input, on a tool call before it runs, on a tool's result, or on the final reply.

Q: Why not just tell the model the rules in the prompt?

Because a prompt is a request, and a model can be persuaded, confused or manipulated. Rules about money, data and irreversible actions must be enforced in the tool's code, where no wording can change them.

Q: What is human-in-the-loop?

A person takes part at defined points: approving an action before it runs, monitoring live, or reviewing afterwards. It is used for costly or irreversible actions.

Q: What is prompt injection?

Text that causes a model to follow instructions its operator did not intend. It may be typed by the user, or hidden in a document or web page the agent reads.

Q: How do you know whether an agent works?

By running it on a fixed set of test tasks in a safe test environment and checking the end result in the data, several times per task, not by reading a few replies.

Answering a limits question with 'I'd write a careful prompt'.

Practise it: Design a Multi-Agent Data Analyst, Design an AI Customer Support Agent

A beginner self-check

What it is

Before moving to the intermediate questions, check that you can do each of these without notes:

define an agent in one sentence and give an example;
draw the loop and label what the model does and what your code does;
explain function calling as four steps;
say three ways an agent differs from RAG;
say where a spending limit belongs, and why;
name four ways a loop can end.

Why it matters

These six are the foundation every later question stands on.

If any is shaky, the intermediate answers will not hold together.
Explaining them to someone outside the field is the best test: if they follow, you understand it.

In the real world: A study pair quizzes each other with these six every morning for a week; both report that the drawing of the loop is what they reached for in the actual interview.

Visual example

Self-check

If a row takes more than a minute, revise the module beside it.

Six things to be able to do without notes

you can	module
Define an agent in one sentence, with an example	1
Draw the loop; label model and code	1, 10
Explain function calling in four steps	3
Give three differences between an agent and RAG	1, 11
Say where a spending limit lives and why	2, 31
Name four ways a loop ends	9

Moving on because the definitions feel familiar. Familiar is not the same as being able to say it aloud in forty seconds.