Sierra interview, decoded.

Two to five weeks is what candidates report, with the in-person stage driving the scheduling. The shape below is the one Sierra published in April 2026; guides that describe a take-home plus a pair-programming round are describing the loop it replaced, so confirm your own sequence with the recruiter rather than preparing for a round that no longer exists.

Sierra publishes no values list, but the post explaining the loop is explicit about what it is trying to see, and each stage exists to test one of them.

• Product thinking: the Plan stage exists so interviewers can watch you decide what is worth building before anyone asks how to build it
• Agency: the post names the question directly, which is whether you pivot when you get stuck
• Judgment under a clock: how you scope two hours of work, and what you choose to leave out
• Technical taste in the code you generate: data model, abstractions, extensibility, the path to production
• Fluency with coding agents, including the ability to say what you delegated and what you checked
• Strength over the absence of weakness: Sierra says its debriefs now ask where a person would thrive rather than whether to hire them

There is no question bank for a loop like this one. What follows is what candidates report being asked around the build, and the technical ground the system design and debugging rounds cover.

Sierra's engineering post of 22 April 2026 describes the loop in the company's own words, which makes this the rare signature round you can prepare for from primary sources rather than from rumour. The old loop was two coding interviews plus algorithms, system design and culture fit; the post says the signal it produced was mostly mechanics, which it lists as typing syntax into an editor, remembering algorithm details and stitching frameworks together, and that when the interviews stopped saying anything, hiring managers leaned harder on referrals and prior experience. The replacement has three stages in one sitting. Below is what each one asks for and where candidates lose them.

• Plan: you choose what to build: A working session where you drive the ideation and the interviewers push on it, on an idea from your own domain. Arrive with two or three ideas you could build in two hours that are worth building: something with one interesting mechanic, in a domain you can speak about with authority. Say who it is for and what you are deliberately leaving out before anyone asks. Treat the interviewers' questions as a spec being negotiated rather than a test being marked.
• Build: two hours, alone, your tools: The interviewer leaves the room. You keep the freedom to pivot and to cut. Sierra tells candidates in advance that cutting scope is fine and that boilerplate such as CRUD and auth should be skipped so the time goes to what is unique, so a candidate who spends forty minutes on a login form has misread the brief. Ship the one mechanic that makes the idea interesting, make it demonstrable, and keep the data model clean enough to talk about. Paul Buchheit's line, quoted in the post: if it is great, it does not have to be good.
• Build: use the agents you actually use: The tooling is yours to choose, which means a candidate who has never driven a coding agent under time pressure is at a real disadvantage. Practise the loop you will run on the day: a short plan in a file the agent can read, small diffs you review rather than accept blind, a test you write yourself for the part that matters. Know where your agent is fast and where it invents things, because you will be asked.
• Review: demo, then defend: You show it working, then the conversation goes to the product flows, the code, and the path to production. The post names what interviewers look at in the code: data model, abstractions, extensibility. Expect to be asked why this shape rather than another one, what breaks at ten times the users, and what you would build next with another day. Name your own hacks before they are found.
• Review: how you used AI along the way: The post says they dig into this, and it is the one question no aggregate prep guide prepares you for. Have a real answer: what you delegated and why, what you checked by hand, where the agent went wrong and how you noticed, what you would have written yourself either way. The weak answer is a shrug; the other weak answer is pretending you did it all by hand.
• What replaced the phone screen: A system design interview, for the stated reason that getting a thing into production scalably is the harder problem. For an agent company that means designing the agent's own machinery: the reasoning loop and its stop conditions, tool permissions and argument validation, state across turns, evaluation and regression suites, and the escalation path to a human. Bring numbers, and say what you would measure once it is live.

Prepare by running the onsite on yourself. Pick a domain you know, give yourself two hours and the agents you use daily, and ship one demonstrable mechanic. Then present it to somebody for fifteen minutes and have them ask why the data model is shaped that way, what breaks first, and what the AI wrote. Do it twice. That is closer to this loop than any number of algorithm problems, and Sierra's post is explicit that the algorithm problems are gone.

The medians come from Levels.fyi's Sierra software engineer page for the San Francisco Bay Area, last updated 30 April 2026 and read on 23 September 2026, from a small sample: treat them as a range to negotiate within rather than a level ladder. The base-only figure is separate and comes from a reading of about 110 public Sierra job postings across 2025 and 2026, which exclude equity. Sierra is private and has raised at a valuation among the highest in AI, so the equity is both the largest part of an offer and the part nobody can price for you. Ask what the grant is worth at the last round's price, over what schedule, and what happens to unvested stock if you leave.

1. 01Read Sierra's own post on the AI-native interview before anything else, then ask the recruiter which rounds your loop has. The published format is newer than most prep material about the company.
2. 02Run a two-hour build on yourself, twice, with the agents you use daily, and present the result to someone who will push on the data model.
3. 03Have three buildable ideas from your own domain ready for the Plan stage, each with one interesting mechanic and an explicit list of what you would skip.
4. 04Prepare the AI question honestly: what you delegate, what you check, where your tools fail you.
5. 05For the system design screen, be able to draw an agent's reasoning loop with its stop conditions, its tool permissions and its evaluation harness, with numbers attached.
6. 06For the debugging pilot, practise reviewing a cross-cutting pull request in a codebase you did not write: run it, read the output, and write the review comments you would leave.

This guide covers Sierra's engineering and research hiring. For management and leadership roles the loop is similar but the bar shifts to people, delivery and strategy, so pair it with the leadership interview prep hub. The bar for your exact role comes from the role-by-role guides, and the prep that actually transfers is spoken, so run a mock interview before the real one.

• 01Sierra, The AI-native interview (22 April 2026)the primary source: the old loop, why it was removed, the Plan, Build and Review stages in the company's own words, the two-hour build with the candidate's own AI tooling, the evaluation criteria shared in advance, the phone screen replaced by system design, and the debugging round in pilot.
• 02Bret Taylor on the AI-native interviewSierra's co-founder introducing the change publicly and asking how others are navigating it.
• 03Four-Leaf, the Sierra interview process and why it deleted the coding round (September 2026)a close reading of the same post, including what prep sites still get wrong about the loop.
• 04TechInterview, Sierra interview guide 2026the agent-shaped system design prompts (a reasoning loop with stop conditions, tool permissions and argument validation, evaluation infrastructure), the practical TypeScript and Python coding style, and the compensation bands quoted by that guide.
• 05Aced (Exponent), Sierra agent engineer interview guidethe agent engineer track, the cross-functional rounds on customer-facing agent behaviour, and candidate-reported timelines.
• 06Levels.fyi, Sierra software engineer, San Francisco Bay Areathe Bay Area median and the highest reported package, last updated 30 April 2026, read 23 September 2026.

Interview processes change. This reflects widely-reported and sourced conditions as of 2026 — confirm specifics with your recruiter, and treat it as a map rather than a guarantee.