
Aki Wijesundara, PhD
AI Founder | Google AI Accelerator Alum

Manu Jayawardana
Exited AI Founder | Co-Founder, Snapdrum
The interview is going well. You have described the architecture, the demo landed, and then someone asks who the agent is when it calls an API. Or what you would look at first when the cost doubles. Or how you stop this becoming forty unowned agents in eighteen months. These are not trick questions. They are the questions, and most candidates have simply never had to answer them out loud.
This resource is those questions, written from the interviewer's side of the table. Ten of them, each with what it is genuinely testing, five points a strong answer contains, the tell that marks a weak one, and the follow-up that comes next.
Inside:
How to answer any of them. Ask one clarifying question first, name the tradeoff out loud, then pick. Say what you would measure, and say how it fails. Four habits that reframe every answer after them.
The build questions. Choosing between the four Microsoft surfaces, grounding against forty thousand documents where half are out of date, designing a tool that writes rather than reads, and who the agent is when it calls an API.
The proof questions. Convincing a security review board that it works, finding where the cost actually went before touching the model, and walking through the hour after an agent gives a confidently wrong answer to an executive.
The judgment questions. Where you draw the seam between what engineering owns and what a business user edits weekly, how you stop an estate of unowned agents accumulating, and how you decide what to adopt when the platform ships three new features a quarter.
The tell on every page. What a weak answer sounds like, in the interviewer's words. Read these first. It is the fastest way to hear what you currently sound like.
How senior candidates sound different, plus the week-before checklist. Five habits that separate a pass from an offer, and eight things to have ready before you walk in.
Free
Written from the interviewer side of the table: ten questions, what each tests, and what loses the room.