The interview simulation: what it is, how it is scored, how to prepare

The Agentic AI Interview at learnagenticpatterns.com/interview is a timed session built on the live challenges. The core set has three questions: the double charge (prompt chaining), the release note with the wrong number (reflection), and the tool called with the wrong argument (tool use). About four minutes a question, twelve for the set. No login to play. You can sign up after the result to keep the score.

Each question runs the same way. The interviewer's opening appears, then the trace builds on a clock, one line per caption: "Here is what came back. Look at the order number." After 25 to 28 seconds the scene pauses: tap the line where it went wrong. Then you say why in your own words, spoken on a phone or typed. One follow up comes back. Then you write the fix, and a real model runs it on every input, including one you have not seen.

Ten points a question, thirty for the set. The tap is worth three: three on the first try, two on the second, one on the third. Three misses is zero, and the page then shows you the line and asks for the cause anyway.

The cause is worth three: zero for nothing, one for the symptom only, two for the real cause, three for the cause plus a change that would fix it. A model grades it against the challenge's own explanation and rubric and never sees the reference fix, so its feedback cannot leak the answer. A keyword scorer runs underneath as a floor: the model can raise your score, never lower it below what your words earned. Silence scores zero.

The follow up is worth one: concrete enough for a senior engineer, so it names what the user sees, what the system does, or what gets measured. The run is worth two: two for a pass, one for needs review, zero for a fail. A pass means the deterministic checks held on every input and the rubric scored at least half. Needs review means the checks held but the rubric was under half, or the checks failed but the rubric was at least two thirds. If the model was unavailable for the run, no points are lost and none earned.

The senior point is the tenth: the thing an experienced engineer says first, matched against your own words across the cause, the follow up and the fix when the session closes. The pattern never reaches the browser. On the double charge it is saying the gate first: plain code that checks the extraction, and a dead letter path for what fails it.

During the scene the clock drives the captions. From the pause it counts up through the tap, the cause, the follow up and the fix. There is no cutoff. The seconds are recorded with your result, and the page says the true thing: your first tap is the one the interviewer remembers.

One budget: twelve model graded answers per challenge per day for one viewer. Past that the keyword scorer alone grades you, so the first attempts of the day are the ones to take seriously.

Read Parts 1 and 2 first. The three questions are lessons 1.1, 1.2 and 2.4, and each ends with the exact sentence to say. The shape from lesson 1.1 is the shape the grader rewards: line number, cause, change. "Line 3. The extraction returned the wrong order id and nothing checked it. I would split extraction from reply and put a string check between them." That is a three on the cause and it contains the senior pattern.

Treat the follow up as a push back. It asks what happens next: what the customer sees, which step may send the email, why ninety out of a hundred is not good enough. Answer with the consequence and the measurement, not with the fix again.

For the fix, write something that runs: on a prompt challenge, a prompt that returns the JSON the checks look for; on the tool challenge, a tool definition whose parameters say what they need. The run uses an input you did not see, so a fix that hard codes the example fails.

The result page shows the score, the breakdown per question, and the senior point text for each. Come back a week later and do it again without the lessons. The number the site tracks is whether you found the bug unaided the second time.