Agent Engineer Interviews in 7 Days: Round by Round, Still Standing at the Third Follow-Up

The Interview Map and Reading a Job Description: What Each Round Judges, and a Skills Matrix From the JD

An interview is not one exam but several, each judged differently. First lay out the interview rounds in China and overseas as one map and see what the interviewer is actually judging in each. Then use three fictional job descriptions to show how to split one into must-haves, nice-to-haves, and differentiators, attach something you can show to each skill, and read off what is left as your gap list. Finally, lay the foundation of the interviewer agent: the model only gives judgments, and every decision lives in code.

Concept Questions, Part 1: A Four-Step Answer and the Family Tree of Agent Frameworks

Put three answers to the same question about ReAct and Plan-and-Execute side by side, one recited, one explaining the mechanism, one explaining the trade-off, and the gap is obvious at a glance. Today you practice a four-step answer: what is being tested, how to break it down, the conclusion, and the follow-up you expect. Then one axis ties ReAct, Plan-and-Execute, Reflexion, LATS, and explicit state graphs into a single family tree: how much feedback the model gets from the environment before each step. The interviewer agent learns to follow up: it asks about whichever point you missed, and once you cover them all it digs into why you did not do it the other way.

Concept Questions, Part 2: The LLM Fundamentals That Change an Agent Engineer's Decisions

Interviews always ask about how LLMs work, but an agent engineer does not need to derive formulas; you need to know which principles change your choices. Today covers only five: why the KV cache means your prefix must stay stable, why tool calls want a low temperature, how to choose between retrieval, fine-tuning, and a long context, how the sycophancy that preference optimization brings affects a judge model, and how batching, speculative decoding, quantization, and mixture of experts change time to first token and cost. Each one ends on what an agent engineer should therefore do. The interviewer agent learns to score: trade-off and evidence points count double, and every quote must really be in the answer.

The Project Deep Dive: Five Layers of Follow-Up, Traces of the Choice, and Where Every Number Came From

The project deep dive carries the most weight of any part of the interview, and it judges exactly one thing: whether you actually did the work. Today the deep dive becomes five layers of follow-up, and you practice leaving traces of the choice at every layer. Then real numbers measured in several courses on this site show how to talk honestly about a course project. The interviewer agent learns to catch claims that cannot be checked: percentages without a denominator, phrases like significantly improved, a story told entirely in we, and numbers nobody can say how they were measured.

Agent System Design: A 45-Minute Time Budget, Back-of-the-Envelope Estimates, and Failure Modes

The most common way to lose a system design round is not drawing the wrong architecture. It is drawing before asking for requirements, not being able to say what the thing costs per month, and not being able to say how it breaks. Today splits 45 minutes into requirements, estimation, architecture, deep dive, and failure modes and evaluation, and drills working out tokens times volume times price and steps times per-step latency on the spot. A pull-request review agent is walked through end to end, and two more prompts come with checklists. The interviewer agent learns to review a design document: are all the stages there, are the estimates right, and how many failure modes does it cover.

Scenario Questions and Coding: A Five-Step Incident Drill and Four Agent-Flavored Coding Problems

Scenario questions ask what you do first when production breaks; coding questions ask whether your code holds up at the edges. First, practice five steps for incidents, symptom, localize, contain, root cause, and prevent recurrence, and walk through five scenarios: looping, hallucination complaints, costs doubling, a tool-timeout avalanche, and retrieval getting worse. Then write four agent-flavored coding problems: merging streamed tool-call chunks, a loop with a budget, retrying only idempotent operations, and trimming history to a budget without splitting a tool call from its result. The interviewer agent learns to run your code: beyond the public test case there are hidden ones, and a failure there becomes the next follow-up.

Behavioral Questions and a Full Mock Interview: Verifiable STAR Stories and a Weakness Report

Behavioral questions ask about the past and judge who you will be: can your results be checked, which part was yours, and what did you change after a failure. Today practices verifiable STAR stories, the line between I and we, failure stories, and the questions you ask at the end. Then the modules from the first six days go into two panels, one China-style and one overseas-style, and each of the three scripted candidates runs through both. The interviewer agent hands back a weakness report where every weakness points to the matching course on this site, and the seven days come together in one sentence.