What the round looks like
One engineering manager candidate was told beforehand to expect an existing codebase, with bugs to fix and features to add. On the day it was features only. The app was full stack with a Python backend, and CRUD operations were missing in the routes, the service layer and the repository layer. The result was checked through the UI and the tests that came with the app, and the candidate was told not to add test cases.
An intern candidate in Sydney got the other version. The backend had a bug in a password feature, and there was an hour to fix it with the assistant. Then came a follow up question on which file the password logic belongs in. An SDE-2 candidate in India got a simple web task in Spring Boot. They were asked to plan the feature with the assistant first, and only then build it. The interviewer graded how the candidate explained the plan, how they used the assistant and how they validated what came back.
The assistant is provided. Most reports put the round in HackerRank's AI IDE. It has a plan mode, which writes a Markdown plan, and an agent mode, which edits files after you approve each step. The interviewer sees your prompts as you type them. One candidate had CoderPad with Claude instead, where every edit had to be accepted by hand, and found it slow.
One report describes a different task, a vague request to design and build a configurable logging system. The candidate was expected to draft the architecture without the assistant, then refine the plan with it without drifting far from the original. After that, implement it end to end, with edge cases. We have seen this once.
What candidates say they were graded on
One invitation said AI use was strongly encouraged. One candidate described the round as pair programming with the assistant. In their words the grading was about how you break tasks down, how you write prompts, how you debug and how you read code. The algorithm mattered much less than in a usual coding round. Another candidate got an internship offer. They said the point was to use the assistant to understand an unfamiliar codebase and fix what was wrong in it.
Two tactics came up. One candidate had the assistant annotate every function in the repository, and the annotations exposed the bug. Another named the failing test and asked the assistant why it failed. That is a smaller and more honest prompt than asking it to fix everything.
The negative report is about guidance, not the task. One Sydney intern candidate was rejected, and said the interviewer gave little direction, both on the coding and on how to use the agent. Our read: the round assumes you already have a way of working with an assistant. It will not teach you one in the hour.
What is confirmed, and what is not
- Confirmed, in more than one report: the starting point is an existing small web app, and the task is missing features or a bug. You plan with the assistant before coding, an interviewer is present, and the assistant is provided rather than your own tools.
- Reported once: the hour as the duration. The existing tests and the UI as the check, with no new tests allowed. The routes, service and repository split, and the follow up about where password logic belongs. Spring Boot as the stack, CoderPad with Claude as the environment, and the logging system task.
- Not known: which levels get the round, how it is weighed against the others, and the rubric.
How we would prepare
- Learn to read a layered app in ten minutes. Where does a request enter, which service method does it reach, what does the repository return, and which rules do the existing endpoints enforce on the way? The missing endpoints have to follow the same path.
- Extend a pattern, don't write a fresh one. If every read endpoint checks board membership before it touches a card, your create endpoint does too. Assistants write endpoints from scratch and skip the checks the neighbours make.
- Treat a failing test as the brief. Before you prompt, read what it sets up and what it asserts. Then ask the assistant a narrow question about that test, not a broad one about the feature.
- Plan out loud before you accept a plan. The interviewer hears your prompts. If the assistant wrote the plan and you did not change it, you cannot defend it when the follow up comes.
- Check the assistant's code against the rest of the app. Does it save the row it changed? Does it keep the order of the others? Does it return the same shape as the other endpoints? The existing tests catch some of that, and the interviewer asks about the rest.
Nyrion's Atlassian section starts with Kanban Board API: Cards, Moves and Passwords, built from these reports. It is a board app in layers, where users and boards already work and cards can only be read. First there is a bug in the password change. It locks users out and refuses some valid passwords, and the follow up is on which layer the fix belongs in. Then you create cards through the routes, service and repository, following the pattern the read endpoints set. Then you move and delete cards without breaking the order of the others. The traps there are the ones code written by an assistant falls into. The layered app, the password bug, the missing CRUD and the tests you cannot change are what candidates describe. The kanban domain, every endpoint and every rule are ours.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
Is the assistant provided, or do I bring my own?
Provided, in every report we have. HackerRank's AI IDE in most of them, with plan and agent modes, and CoderPad with Claude in one. We have not seen a report of candidates using their own tools in this round.
How long is it?
One report says an hour. The others do not say, and Atlassian has not published it.
What language is the app in?
Python in one report, Spring Boot in another. Nothing suggests you get to choose.
Can I add tests?
One candidate was told not to. The result is checked with the existing tests, and with the UI where there is one.
What does the interviewer do?
They are present and they see your prompts. In one report they asked a follow up about where the password logic should live. In another they gave little guidance and the candidate was rejected, so don't wait for direction.
Does Atlassian say anything about this round?
No. The engineering interview guide on its careers site lists data structures, code design and system design, and does not mention AI. Everything on this page comes from candidates.
Practice problems tagged for Atlassian
Browse the catalog and open the Atlassian section. No login required to look; sign in when you want to solve one.