The Interview Loops at the Top Labs

Every FDE interview loop in 2026 is screening for the same small set of behaviours, dressed up in three different ways. This lesson is not a cheat sheet. It is the shape of the loops as they stood when this course was written, the signals interviewers read for in each round, and the rounds candidates most reliably fail — so you can practise the right ones before you sit down.

One caveat first. Loops change, sometimes every quarter. Treat what follows as a map, not a timetable, and verify the specifics at the careers page of the company you are interviewing with before your recruiter call.

The three loops at a glance

The common shape under all three is coding, a case study, a customer-facing round, and a judgement conversation. The surface differs. Anthropic leans on its own agentic surface — know how to author an MCP server and when to reach for a sub-agent, not because they will ask you to recite the protocol but because that is the vocabulary your case study will be assessed in. OpenAI leans on evals, because that is the discipline their JDs most insist on and the one candidates most reliably underweight. Palantir leans on ambiguity, because the day job is built out of ambiguity and the loop is built to screen for it.

What interviewers read for

Three signals show up, with different weights, in every round:

• Can you clarify ambiguity? A prompt in an FDE loop is rarely a complete problem. Candidates who ask a handful of sharp questions before proposing anything score higher than candidates who make the problem whole by assuming. The specific questions do not need to be clever — "what does this look like today?" and "who decides when this is done?" are the two that do most of the work in a real customer room and in an interview room.
• Can you cut scope? The problems are sized to be unfinishable. The candidates who try to do everything in 45 minutes end up with a diagram that is correct in outline and useless in detail. The candidates who say "I am going to carve off this slice and leave these two for a follow-up" are read as FDE-shaped because that is exactly the move the job asks for on week one of a POC.
• Can you say you would measure it? The strongest tell in an applied-AI round is the candidate who, asked which model or which retrieval strategy or which chunking size, answers not with a specific choice but with a specific test. "I'd try both and look at this number" reads as mature. "It's probably X" reads as salesy.

The rounds where these three show up most are the case study and the customer simulation. Coding rounds still matter — you are expected to be competent — but they are not where the role is decided.

The rounds candidates most reliably fail

Evals (OpenAI, implicitly at Anthropic). Candidates arrive strong on models, prompts and retrieval and reach the eval round with the one-chapter mental model that most curricula have given them. Expect a scenario in which an eval already exists and is lying — a judge that scores confident wrong answers as correct, a dataset that overlaps with training data, an online metric that does not match the offline one. The move is not to name a specific fix. It is to name the failure mode and propose a cheap test that distinguishes it from the alternatives.

The ambiguous customer problem (Palantir). A paragraph of information, most of a workflow missing, a stakeholder with a strong opinion that is wrong, a constraint that is only mentioned once. Candidates fail by pretending the missing information is present and designing around it. Candidates succeed by naming what is missing, asking for it, and — if the interviewer refuses to supply it — stating what they would decide in each of the two or three possible worlds.

The customer simulation (Anthropic, increasingly OpenAI). A live roleplay in which the interviewer plays a difficult customer and the candidate plays themselves. Candidates fail by defending, agreeing with everything, or reaching for a demo. Candidates succeed by asking questions, summarising what they heard back to the "customer", and being willing to say "that is outside what I would commit to" on camera.

A session you can practise against

customer-sim

2. >You typed: Before I answer, a few questions. How many do you draft a week, and what portion of them use the same template? Who approves each one today, and what does 'wrong' look like in a draft?

3. WhyFour questions, under 30 seconds, and the interviewer has already read the signal. You are not treating 'build an agent that writes NDAs' as a complete problem.

5. >You typed: So the risky failure is a silent dropped clause, which means the eval needs to measure clause coverage and not just text similarity. Can you give me 50 past NDAs and the counterparty terms as a spreadsheet?

6. WhyYou have reached for an eval before an architecture. That is the sentence Anthropic interviewers remember.

What to take away

The three loops screen for the same small set of behaviours through different surfaces: Anthropic through its agentic stack, OpenAI through evals, Palantir through ambiguity. The signals that win rounds are the ability to clarify a problem rather than complete it with guesses, the willingness to cut scope rather than attempt everything, and the habit of saying what you would measure rather than naming the thing you are not sure about. The rounds candidates most reliably fail are the eval round, the ambiguous customer problem, and the customer simulation — because each one punishes the behaviour the rest of a career optimises for. Practise the rounds that will decide the loop, not the ones that are easiest to practise, and verify the loop at the careers page before you sit down.

Next: a worked case study walking the whole arc of an FDE engagement from discovery to a day-14 incident and back.