Why Coinbase's version is different
Meta hands you an algorithm and lets the assistant write it. Amazon hands you a broken repo and an assistant that will not. Stripe hands you an easy spec and grades how you delegate. Coinbase did something else: it took the question it already trusted, the transaction filter and pagination endpoint that has nine candidate reports of its own, and reframed it as maintenance. You no longer build the pagination. You inherit it, with bugs, logs and a pull request, and you work on it with an assistant while someone watches.
Our read is that this is the most realistic of the AI rounds we have covered, and the hardest to fake. Two of the four parts have no code deliverable at all. You design, you diagnose, you review, and you do it out loud. The assistant cannot do any of that for you, which is presumably the point.
What the round actually is
The IC5 candidate who reported it in July 2026 described the whole script. The service paginates transactions that carry a timestamp, an amount and an id, in memory, sorted by timestamp. Then four parts:
- Debug. The sort-by-timestamp pagination has problems: an invalid cursor format, duplicate timestamps that need an id tie-break so rows are not skipped or repeated, and a cursor change that has to stay backward compatible with clients still sending the old one.
- Design. Add sorting by amount. The interviewer wants a trade-off analysis across four approaches: offset versus cursor pagination, a cursor that does or does not embed the sort key, and a cursor that is hashed or opaque versus plain.
- Diagnose. From logs, work out why a particular request keeps timing out (there is no index on the amount column) and propose improvements, including options that do not add an index.
- Review. Another engineer's code, on a screen share, with the assistant able to see the task description and the open files. Say what you would accept, change or reject.
A May report from a different loop describes the same four shapes in the same order (find a bug, propose a feature, review a PR, debug a production scenario) over what that candidate called a mock object-oriented database with cursor queries. An April comment calls the task "the well-worn pagination question, asked in a much more open way", with the second part being "there are many alternatives, which do you think is better and why". Three reports, one task.
The question it grew from
If you want to understand the round, read the nine reports of the classic version. Candidates are asked to build a transaction query over an in-memory list: filter by user, currency, time range and amount range, then paginate. The interviewer does not say what the API should look like. The round is decided in the first five minutes by whether you raise offset versus cursor pagination unprompted, tie-break the cursor on the timestamp and the id, and can say in sixty seconds why cursor is the production answer for a history that people scroll rather than jump around. Several candidates failed it for typing before negotiating the API.
The AI round is that same conversation with the code already written. The duplicate timestamp bug is the tie-break you were supposed to volunteer. The four cursor designs are the trade-off you were supposed to argue. The timeout is what happens when someone sorts on a column without an index. If you can do the classic version well, the AI version is a change of format, not of substance.
What is confirmed, and what is not
- Confirmed: an onsite round with an in-IDE assistant and a mandatory screen share; four parts (bug, feature design, production diagnosis, code review); a cursor-paginated transaction service as the codebase; the interviewer recording how you prompt and validate; dense follow-up questions and time pressure. Each is in two or three independent reports.
- Reported once: the exact bug list (cursor format, duplicate timestamps, backward compatibility); the four cursor designs; the missing index on amount; the assistant seeing the task description and open files.
- Not known: the rubric, which model powers the assistant, and whether every loop gets this round. The round is new enough that interviewer calibration varies.
How we would prepare
- Write cursor pagination once, properly. Filter, sort by (time, id), position by the last row seen, take limit plus one to know whether there is more. Then break it the way the round does: drop the id from the cursor and watch rows vanish at a duplicate timestamp.
- Rehearse the four-way cursor trade-off. Offset, id-only, sort key plus id, signed. Sixty seconds each: what it costs per page, what happens under inserts, what a client can forge. This is the part most directly graded.
- Practise reading a log for a cause. Latency that does not depend on the page, a queue that grows, timeouts that hit unrelated requests. Say the cause, then give two fixes: the index, and what you would do if you could not add one.
- Review a diff out loud. Thirty seconds per change: correct because, should change because, fine but add a test. Most people either rubber-stamp or nitpick. Neither lands.
- Read first, prompt second. Candidates who pasted the file and asked for a fix got vague answers. Skim, form a hypothesis, ask the assistant to confirm a specific thing.
Nyrion's Coinbase section starts with Transaction History: Cursor Pagination, built from these reports. A small service with the three reported bugs, a production log to diagnose, a PR with planted defects to review, and the amount sort to add with tamper-evident cursors. Each stage ends with the interviewer's questions: why rows went missing, which cursor design and why, what the log told you, what you would have written on the PR. Our assistant writes code, which Coinbase's mostly will not, so practise asking it to confirm what you already suspect.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
Is this a coding round or a talking round?
Mostly talking, with code on screen. One candidate described it as almost pure conversation over a screen share. You read an existing codebase, prompt the assistant, explain what it told you in your own words, and the interviewer records how you prompted and how you validated the output. Expect to type less than in the classic coding rounds and to explain more.
What is the codebase?
In the fully described report, a transaction history service with in-memory cursor pagination over records that carry a timestamp, an amount and an id. That is the same system behind Coinbase's transaction filter and pagination coding question, which has nine candidate reports of its own. One earlier report describes it as a mock object-oriented database with cursor queries, which is the same thing seen from a different angle.
What are the four parts?
Find and fix a bug in the existing sort-by-timestamp pagination (an invalid cursor format, duplicate timestamps that need an id tie-break, keeping the cursor change backward compatible). Design sort-by-amount support and argue the trade-offs between four cursor designs. Diagnose from logs why one request keeps timing out and propose fixes, including ones that avoid adding an index. Review another engineer's code on a screen share. Every part comes with a stream of small follow-up questions.
Does the AI assistant actually help?
Reports say little about it beyond what it sees: the task description and your open files. One reader asked whether it only answers basic questions and you still produce the solution yourself; nobody who sat the round answered. Treat it as a fast way to confirm a hypothesis about code you have already read, not as the engineer.
Where does it sit in the loop?
It is one of the four virtual onsite rounds, listed in the schedule as AI Enhanced Pilot or Domain. The others in the reports: a classic multi-level coding round (block mining, crypto order management, the banking system), system design (the Explore price page, the trading flow), and a hiring manager round that now asks how you use AI day to day. Before the onsite there is a cognitive test, a personality questionnaire and a CodeSignal OA, which on some loops is itself AI-assisted (a Flappy Bird strategy function).
How is it graded?
Coinbase has published nothing. The one candidate who shared feedback was told well done on the round and was still rejected on the loop overall, so a good AI round does not carry a weak classic coding round. The reports agree the interviewer tracks three things: whether you read before you prompt, whether you verify what the assistant says, and whether you can turn its explanation into your own.
Practice problems tagged for Coinbase
Browse the catalog and open the Coinbase section. No login required to look; sign in when you want to solve one.