Sierra's onsite gives you 2 hours, any coding agent and $150 of tokens

How Sierra's 210-minute AI-native round works: 30 minutes choosing a project with the interviewer, 120 building it alone with coding agents, then a demo and a code review you lead.

Wood chisels laid out on a pale workbench marked with pencil lines
Photo: Dominik Scythe on Unsplash
  • You choose what to build. The first 30 minutes are spent with the interviewer picking one of your ideas and cutting it down to something you can demo.
  • Then you build it alone for 2 hours, on your own laptop, with Claude Code, Codex, Cursor or whatever you normally use. Sierra pays back up to $150 of tokens.
  • The last hour decides it: a short demo, then 45 minutes where you lead a code review and explain how you would take the prototype to production.
  • They want depth, not a full product: cut the boilerplate, make one part of the code excellent, and change course early if the idea isn’t working.

Where the round sits in Sierra’s onsite

Sierra’s onsite for platform engineers has three parts. The AI-native round tests whether you can take something from nothing to a working demo. A code ownership round tests whether you can work in code someone else wrote. A conversation with the hiring manager goes through your past work, with a focus on agency and how you worked with others.

The AI-native round is the long one: 210 minutes, about 3½ hours: plan, build, then an hour of demo and review.

PhaseMinutesInterviewer
Plan30With you
Build120Out of the room
Demo10With you
Code review and path to production45With you
Questions5With you

Plan (30 minutes): pick the project together

You come with ideas; the interviewer helps you choose one and scope it. Sierra suggests ideas from your own experience: an interesting project you worked on, a small version of a feature you like in a product you use, or a tool you wish existed. You don’t write code yet.

This phase is graded too. Sierra says it is looking for speed, creativity and collaboration: whether you can settle on something quickly, whether the idea shows what you are strongest at, and how you react when the interviewer pushes back or changes something. Expect them to.

A good way to spend the 30 minutes: pitch two ideas in a sentence each, agree on one, write down the three or four steps the demo will show, then name the part you will build properly and the parts you will fake (sign-in, settings, admin screens).

Build (2 hours): alone, with your coding agents

The interviewer leaves. Their calendar stays blocked, so you can reach them, but the time is yours. You work on your own laptop with the AI coding tools you prefer, and nothing about how you use them is restricted.

Sierra spells out what it expects from these 2 hours:

  • Use your judgment, and pivot if you need to. The build is deliberately open. If the first idea turns out weak or doesn’t work, changing it is fine. Sierra suggests checking in with yourself about an hour in.
  • Work backward from the demo. The demo comes first after the build, so make sure the main workflow runs end to end, and run through it once before the interviewer returns.
  • Depth over breadth. Cutting scope is fine and expected. Skip auth and CRUD screens and spend the time on what makes your idea interesting.
  • Show technical craft somewhere. You can let the agent write the front end, as long as another part, such as the back end or the API design, shows your best engineering. You will review that code together.

Demo and review (60 minutes): you lead it

The interviewer comes back and the last hour runs in three parts.

  • Demo, 10 minutes. Show the workflow you agreed on. A demo that breaks on the first click costs you more than a missing feature.
  • Code review and path to production, 45 minutes. You drive this, with questions and input from the interviewer. Walk through the code you are proudest of, say what the agent wrote and what you changed, and be honest about what you would not ship.
  • Your questions about Sierra, 5 minutes.

The path to production works like a system design interview about your own project. You don’t need to have built a scalable system in the prototype, but you should have thought about how to get there: how the data model and storage change with real usage, what fails and how you retry, auth and permissions, monitoring, testing (and evaluation, if the app calls a model), cost, and what you would build first.

How to prepare

  1. Set up your laptop and prove it works. Your coding agent, and a starter in a stack you know well (Sierra’s example is a Next.js app with a Flask back end and a local SQLite database). Build one small toy app end to end with it beforehand. That app is only practice; you won’t use it in the interview.
  2. Bring two or three ideas from your own domain, where you know what good looks like. For each, know the demo steps and the one hard part you would build properly.
  3. Practice a timed build. Give yourself 2 hours with your agent on one of the ideas, and stop at the hour to decide whether to cut scope.
  4. Practice reviewing code the agent wrote. You will explain it line by line, so read what it produces and rewrite the parts you couldn’t defend.
  5. Prepare the production story for each idea the way you would for a design round.

We wrote the round up as a practice question, with clarifying questions, what a strong answer covers and the follow-ups to expect: AI-native build: scope an app, build it in 2 hours with coding agents, then demo it.

The next round: code ownership

Where the AI-native round goes from zero to one, code ownership goes from one to ten. You get one of two versions and are told which ahead of time: take over a teammate’s pull request and get it to a shippable state, or add a feature to a small existing app. Either way you run the app locally and talk through scope and edge cases before you change anything. The rules on AI differ:

  • DebuggingAI tools only to navigate, explain and make narrow changes, never to review the PR or find the issues.
  • Feature developmentFully allowed.

Two practice versions of the debugging round: Retention agent: fix three bugs in a discount offer engine and Take over a teammate's PR: make live streaming and cancel actually work.

See every reported question for your loop, by round, with when it was last asked.

Pick your company

What the labs asked this week

A weekly email of what OpenAI, Anthropic and the other top AI labs asked engineers, round by round.