The short version
Plan on about eight and a half hours across one week. The schedule, preparation, and rules appear below.
- ~8½ hyour time across one week, including the survey and the build
- 4 hthe build, scheduled by you, on your own tools
- ≤ 5 daysbusiness days from your screen call to a decision in writing
- 1 weekpaid, onsite with the team, before the offer
- 0puzzles or whiteboard trivia. AI tools open in every round
What we're looking for
You try new technology early, take it apart, and decide where it earns a place in your work.
You choose a practical path through ambiguity, explain the trade, and keep moving.
You find the largest useful lever and leave low-value work unbuilt.
You can explain what you ship all the way down.
You model the whole system, reason from first principles, and defend architecture choices with the fundamentals beneath them.
We ask about demanding work, setbacks, and recoveries. We never ask about your personal life.
Coding agents are part of how you build. You adapt your tools to the work and own every line they produce.
You can run a product or direction end to end: define its purpose, choose its scope, and own the result.
We assess your evidence regardless of school, resume length, follower count, accent, or public GitHub presence. Share any evidence you can legally show.
The survey's agent-practice section is where we first see AI-native engineering and system thinking in your own work: how you command agents at scale and the system-level decisions behind it.
The process
- 01About 35 minutes
asyncA surveyAbout five minutes of questions, then about 30 minutes on your agent practice. You complete it on your own time, before any call.
The questions: your work eligibility for the role's location; whether you can work onsite five days a week, your earliest start date, and your notice period; and an honest question about pace. Days often run past 10 hours, and we occasionally work weekends for urgent deadlines, releases, or the occasional weekend hackathon. Tell us whether that pace works for you; you don't need to explain why. The role's compensation range is shown in the survey, and we ask whether it works for you. We also ask about your timeline, including any other offers or deadlines. That never counts against you.
Your work: your GitHub profile link, which is required, plus any other links to your work, such as a site, X, a portfolio or a blog, which are optional. If GitHub doesn't show your best work, name one personal project you can show live on the screen call.
Your agent practice: CodeBurn and AgentsView full-history reports from your main development machine, and a narrated 10–20 minute screen recording that walks through both and through the most agentic, challenging project you've worked on: how you commanded agents at scale. All three are required. The setup is under How to prepare.
Last, two short answers about what's on your radar (an open-source project you find interesting right now, and a recently launched product you discovered), and three one-line nominations: a demanding project that went wrong and you recovered, something you built that nobody asked for, and something you taught yourself and then used. Names only, not stories. We ask for one of the stories later, live.
- 0230 minutes
hands-onA screen callShare your screen from the machine you use for coding and show us your working setup. Run one or two short prompts we provide in your own coding agent and project, then walk us through a highlight you are proud of. We ask a few questions and leave time for yours.
If you advance, we send one email within the hour with every remaining session booked. We mark sessions after the next step as provisional.
- 034 hours
within 2 daysThe buildWork on a realistic problem using our actual tools, at a time you choose. AI is open. One engineer provides the same bounded support a colleague would. We change a condition partway through. You decide what to do first, what to skip, what to verify, and what to ship. Finish with a handoff note a colleague could continue from, including about 150 words on the approach you chose, the one you rejected, and the failure you would worry about first. We send the logistics brief before you begin.
- 042h40
two callsTwo interviewsThe first is 100 minutes with a break, and it is almost entirely systems: 20 minutes on the build you did, 45 designing a system we intend to build and have not built yet, and 30 reading an architecture and saying what is wrong with it. You can use any tool. Expect us to change a constraint on your design twice and ask how it responds, and to ask who gets paged when it fails at 3am.
The second is 60 minutes on product and taste, with two people: 45 on a product or direction you owned end to end, then 15 on a set of design choices and why you would pick one. Seventy-five minutes if you are interviewing as a Design Engineer.
Our CTO also meets you for an unscored coffee, in person when you are in San Francisco and over video elsewhere. The format does not affect the decision.
- 05≤ 5 business days
from the screenThe decisionYou receive it in writing within five business days of the screen call, either way. A yes includes an invitation to the paid week.
- 065 working days
paidThe paid week onsiteSpend five paid working days at the role's office. We send the rate and travel terms in writing before you commit. During the week, you work on one real problem with the team and meet our CEO. You demo and receive the offer decision on the last day.
Tell us in the survey. We compress steps 2 through 5 to 72 hours and can shorten the onsite week.
How to prepare
The survey's agent-practice section takes about 30 minutes. The screen call takes about ten minutes to prepare.
Before the survey
Run both reports on your main development machine. If you develop across several machines, run them on each and include every file. You need Node.js (for npx) and Python with pip, or uv (for uvx), on that machine. Both tools run locally and read the session logs your agents already keep. Nothing is uploaded unless you send it.
Leave confidential projects, such as your employer's, your lab's or advisor's, or unpublished work, out of both reports with the exclude flags. We would rather see less than have you share something you shouldn't. CodeBurn lists each project's full path, so on a shared server keep to your own folder with --project ~ or drop a folder with --exclude /path. The AgentsView file does not list project names, but it does repeat any name you pass to --exclude-project, so only use it for names you're fine sharing.
Both tools read only the agent history in the home folder of the account that runs them (for example ~/.claude and ~/.codex). They don't read the folder you run them from, or other users' folders, even if you can open them. Run them as yourself on the machine where the agent actually ran; for SSH or remote development, that's the server. Check first with npx codeburn doctor: the paths it reads should all be under your home folder. If several people share one login, don't run them; tell us instead.
# CodeBurn full-history report npx codeburn report -p lifetime --format json > codeburn-$(hostname).txt # leave a project out: --exclude <project-name> (repeat for each) # AgentsView full-history report pip install agentsview # or run it without installing: uvx agentsview agentsview stats --format json --since 2020-01-01 > agentsview-$(hostname).txt # leave a project out: --exclude-project <project-name> # Several machines: run both on each, then combine into one file per tool (macOS/Linux) for f in codeburn-*.txt; do printf '{"file":"%s","report":' "$f"; cat "$f"; printf '}\n'; done > all-machines-codeburn.txt for f in agentsview-*.txt; do printf '{"file":"%s","report":' "$f"; cat "$f"; printf '}\n'; done > all-machines-agentsview.txt
The upload box takes .txt, not .json or .zip. The file is still JSON inside; only the name ends in .txt.
In PowerShell, run npx codeburn report -p lifetime --format json | Out-File -Encoding utf8 codeburn-$env:COMPUTERNAME.txt, and the same pattern for AgentsView: agentsview stats --format json --since 2020-01-01 | Out-File -Encoding utf8 agentsview-$env:COMPUTERNAME.txt. Any file name is fine.
Run both reports on each machine, then combine them into one file per tool with the loops above, and upload all-machines-codeburn.txt and all-machines-agentsview.txt. If that's a hassle or the upload still refuses your file, upload your main machine's file and reply to our email with the rest attached.
Record a narrated walkthrough of roughly 10–20 minutes on your main development machine. Cover three things: run npx codeburn and open the Lifetime view (press 6), then say which projects, models and kinds of tasks dominate and what that says about how you work; run agentsview serve and open http://127.0.0.1:8080 to walk through your usage and activity; then the most agentic, challenging project you've worked on, and how you commanded agents at scale: how you broke the work down, ran agents in parallel, kept them on track and verified their output, what broke, and the system-level decisions you made. If you use several machines, show each machine's views. The Lifetime view and the AgentsView dashboard can list every project, including ones you left out of the reports. If a confidential project appears, crop, blur or cut that part of the recording. You never need to name it. Use the same filters on screen, and stay on summary views (individual sessions show your prompts).
Upload it to Loom, Google Drive, YouTube (unlisted) or similar, and set it so anyone with the link can view it. Never show employer code or confidential data. Describe a work project without showing its code.
Before the screen call
Join from the computer you normally use for coding and prepare to share your screen. If you develop through SSH, a devcontainer, or a cloud workstation, open the terminal you use. Close confidential material first. Install Node.js 22.13 or later in that environment.
Open a project you can legally show, its repository if public, and one commit whose story you remember. We give private work the same consideration. Tell us and we will use an equivalent path.
Set up your coding agent, such as Claude Code, Codex, or an equivalent, with the configuration and plugins you use day to day.
Show us the setup you use for daily work. Leave your dotfiles and terminal as they are. Do not create a repository or rehearse a demo for the interview. We want to see your working environment and hear an honest account of private work.
Your rights, our rules
- We never ask your current or past salary.
- We assess communication effectiveness. We never score accent or fluency.
- We ask for your GitHub profile link. A sparse or mostly private profile is fine, and a social presence is optional. Any work you can legally show carries the same weight.
- On the screen call, you may skip private material, exclude any project from view, and opt out of recording the call without penalty.
- The survey's agent-practice section, both reports and the recording, is required. You choose what the recording shows, and you can exclude any project from it and from both reports.
- The pace question asks only whether the pace works for you. You never need to explain why.
- Never show employer code or confidential material. If it appears, we stop the screen share and remove it from our notes.
- AI is open in every round. You remain responsible for the result.
- The loop runs in English.
- The onsite week is paid work. We put the rate in writing and cover travel.
- Every candidate receives a response within the stated timeline. A no arrives by email, not silence.
Questions we get
These roles are onsite five days a week at the office named in the posting, San Francisco or Shanghai.
Yes. We want to see how you work with agents and how you own the result.
A week of real work lets both sides assess the team, the problem, and the working relationship. You are paid, you see the role up close, and you can leave with a reference either way.
We welcome it and do not require it.
This role is intense. Days often run past 10 hours, and we occasionally work weekends for urgent deadlines, releases, or the occasional weekend hackathon. We ask in the survey so expectations are clear. Just tell us whether the pace works for you.
Both run locally on your machine and read the session logs your agents already keep. Nothing leaves your machine unless you send it, and you can exclude any project with the exclude flags.
Run both tools as yourself on the machine where your agent actually ran; for SSH or remote development, that's the server. They read only your own home folder (for example ~/.claude and ~/.codex), not other users' folders. Check with npx codeburn doctor first. If several people share one login, don't run them; tell us instead.
They are short live tasks in your own agent and project. They require no preparation.
Reply to any email a member of our team sends you. We answer in writing the same day.
HellYeah builds AI agents that manage real advertising budgets. Our interview uses the same tools and decisions the job requires.
Back to top