What the round looks like
You share your screen and open whatever you use. The task arrives small: write a client that connects to the warehouse API, starts a query, and reads back its status. The assistant produces it quickly; one candidate said the code itself came in a single pass. Then the interviewer asks about the parts the task left out: a result too large to hold, where the credentials live, how you would run many statements at once.
The failure mode in the reports is not the code. One candidate lost most of the hour to environment and dependency trouble and connected with ten minutes left. Another finished and still could not say what the interviewer had been looking for, writing that the hard part was not the problem, which was too simple for the assistant, but something else.
What the round seems to measure
A line attributed to Datadog's own guidelines, quoted by a candidate in August 2026, says it: they do not want the AI to work for you, they want you to tell it what to do. Read with the follow-ups, the round is a production-sense interview with an assistant in the room. The query client is the warm-up. The questions that matter are what you do when the result does not fit in memory, how a secret gets into the process without landing in the code, and how a batch stays inside the service's limits while still finishing in reasonable time.
Our read: treat the half hour as two problems. The first is logistics, and it is on you to arrive with the environment working. The second is the conversation, and the assistant cannot have it for you.
What is confirmed, and what is not
- Confirmed: a dedicated AI coding round since spring 2026; any AI tool, screen shared, the interviewer watching the collaboration; the warehouse query client as the task; the follow-ups on large results, secrets and batching. Each appears in more than one report.
- Reported once: Claude Code named as allowed; the thirty minutes lost to dependencies; the guidelines quote.
- Not known: how the round is scored, whether the task varies by team, and what the published guidelines say beyond the one quoted line.
How we would prepare
- Have a working environment before the call. The SDK installed, a credential path that is not a hard-coded string, a way to run a script. The reported time sink was none of the interesting parts.
- Know the shape of an asynchronous query API. Submit, poll a status, fetch pages. Say the loop out loud before the assistant writes it, and check what it does on failure and on timeout.
- Answer the three follow-ups before they come. Stream or page a large result instead of holding it; read secrets from the environment or a manager; batch under a concurrency limit with retries for transient failures.
- Narrate what you tell the assistant and why. The screen is shared so the interviewer can see the prompts. A vague prompt and an unread patch are what the guidelines line is about.
Nyrion's Datadog section starts with Warehouse Query Client, built from these reports. A fake warehouse stands in for the service: statements run asynchronously, status is polled, results come in pages, there is a concurrency limit and the occasional transient failure. You write the client that runs one statement, then streams a large result page by page, then runs a batch under the limit with retries, which is the follow-up list made concrete. The task and the follow-ups are what candidates describe; the fake service, the client's interface and the batching rules are ours.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
What is Datadog's AI coding round?
A round introduced in spring 2026 for software and machine-learning engineers in which you use any AI coding tool you want, local agents such as Claude Code included, share your screen, and the interviewer watches how you work with the assistant. Three candidate reports from April to June 2026, and a later comment that Datadog now publishes guidelines for its AI-assisted rounds.
What was the task?
Connect to the Snowflake API, submit a query and check its status: a small query client. Two reports describe it the same way. The follow-up questions were the real content: what if the result or the data volume is too large, how would you handle secrets, how would you support batching.
Is the problem hard?
No, and the candidates say so. One wrote that the assistant solved it in one go and that the hard part of the round was something else they could not name. Another spent more than thirty minutes on environment and dependency problems and only connected in the last ten.
What does Datadog say it wants?
A comment from August 2026 quotes a line attributed to the company's guidelines: they do not want the AI to work for you, they want you to tell it what to do. We have not seen the guidelines themselves.
Which assistant should I bring?
Your own, whatever you are fast with. The reports name Claude Code as allowed. Whatever you bring, have the environment, the SDK and your credentials handling working before the call; that is where the reported time went.
Practice problems tagged for Datadog
Browse the catalog and open the Datadog section. No login required to look; sign in when you want to solve one.