AI-Assisted Coding Interviews

Meta calls it AI-native coding. Reports describe two such rounds in the onsite, next to one or two AI-enabled system design rounds and a behavioral round. The phone screen is still two LeetCode-style problems without AI.

You get a folder of existing code with tests already written, and a chat panel where you can pick the model. Candidates name Claude Opus and GPT models. One candidate notes the available models lag public releases by a month or two. The work comes as a single problem in several parts, and after each part the interviewer asks you to explain the algorithm the AI wrote.

Reported questions: a maze solver in four or five parts (fix a display bug, fix a search that never terminates, add tiles that force a direction, add keys and doors, which means the visited state has to include the keys you hold, then a bomb that clears walls); the largest set of words with no repeated letters, first on about a dozen words and then on tens of thousands; friend recommendation, from fixing a validity check to designing a strategy that runs on about 100,000 users and 1.5 million friendships; a card game where you repeatedly pick three cards summing to 15 and must benchmark your strategy over a thousand games; a pathfinding problem; and speeding up a brute-force search for one string inside another.

The pattern that sinks people: the last part is a scale test. One candidate's solution passed on the small map, then the interviewer opened a very large input and it timed out, and the rest of the round went to arguing with the assistant. Another measured the sandbox at about 1.5 CPU cores and 2 GB of memory, and warns that a modern assistant will run the big test on its own unless told not to.

The rules here are the strictest reported. Both the recruiter and the interviewer told one candidate the algorithm must be theirs and the AI may only help implement it.

The reported question is an LRU cache in four parts: plain LRU, then a TTL, a third part the candidate did not recall, then thread safety. The candidate designed the hashmap plus doubly linked list, then spent the round correcting the assistant, which forgot to remove evicted keys from the map and invented variables. They finished two of four parts, and the feedback named this round.

A commenter who said they were a LinkedIn interviewer described the training: they look for prompts that show you digested the problem and chose the data structure and algorithm yourself, then used the AI to write more code than a normal round allows and to help review it. A stronger model is not the point.

Rippling's phone screen is now reported as an AI coding round of about 60 minutes. One candidate was told it could not be opted out of. You set up your own IDE and assistant beforehand. In one report you copy a single starter file into it and paste the result back into HackerRank to run the tests.

The questions are the ones already in Rippling's catalog: the expense rules engine, the task manager, the delivery driver payment system, and the card-hand problem. A rate limiter is also reported.

The reports disagree on pace, and that is the lesson. The candidate who passed kept the AI closed for about twenty minutes, talking the design through with the interviewer and typing the structure by hand, and only then asked it for function bodies. A candidate who solved a rate limiter with Cursor in a few minutes was left with an interviewer who had nothing more to ask, and was rejected the next day. Interviewers ask about complexity and why you designed it that way.

An entry-level candidate had AI coding, behavioral and system design back to back in one room. The AI coding task was a task queue: a worker pulls jobs, retries with exponential backoff (the interviewer wanted jitter and a cap), and sends jobs past the retry limit to a dead-letter queue. A follow-up added job priority, and the discussion was one priority queue against one queue per level, including how the client input differs.

The interviewer said afterwards what they grade: how you ask the AI, how you write prompts, and how you judge the tradeoffs. Even when the candidate already had a solution, the interviewer pushed them to ask the AI for alternatives. The built-in assistant took minutes per answer. Snowflake also runs a separate 20-to-30-minute screen where an AI asks behavioral questions.

The Power Day is three rounds: behavioral, a case study, and agentic coding. The agentic round gives you a full-stack app (React and Vite in front, Python and FastAPI behind). You find and fix its bugs, then build a new feature, using Claude to locate and explain code while showing you understand each change.

The case study in the same report is a chatbot: the business benefit first, then explaining and improving a piece of code, then talking through how new requirements would change it. No code is written in that round.

A new onsite round in three parts on an expense and budget-report codebase too large to read by hand. Part one: monthly reports are wrong because some expenses are filed in a later month than they happened; debug it. Part two: add a report feature until the tests pass. Part three: brainstorm what production would need and have the AI build it. The follow-up picks functions the AI wrote and asks whether you understand them and how you would optimize them.

Two reported tasks. One gives you an API document and an agent starter script, plus an LLM key and an API key, and asks you to build a chat agent that saves passenger details, searches flights and books one through tool calls.

The other is a feature-flag evaluator: flag definitions, ten thousand users and a file of expected decisions for 200 of them. The traps are in the data: a disabled flag returns its default, regular expressions only match from the start, some rules roll out to a percentage of users, two rules can match the same user, and some user fields were blanked after the expected answers were generated. The candidate's advice is to reverse-engineer the expected file before writing the engine.

60 minutes to build an internal go-link tool for about 100 users. Part one: create a short link and redirect on visit, in memory. Part two: login, link ownership (anyone can see links, only owners can delete), and analytics on the most-visited links and links per user. The candidate says the interviewer grades how you scope the work, direct the AI and verify its output. The assistant was throttled and slow, they ran out of time, and they were rejected.

Five tasks in 90 minutes, with a live preview of the web app. Two tasks are a chat with an AI that hands you the requirements, two are building them with Claude Code, and one is a bug fix without AI. The first build adds search and filtering. The second is much longer: a notification service with in-app availability alerts and due-date reminders, with idempotency and async edge cases. A score of 650 out of 1000 was a rejection.

Ramp gives you a repository and a $50 stipend to set up whatever AI tools you want, then asks you to build a matching system in any language. If you finish the core, you pick bonus tasks.

From 2026, several Stripe reports describe the programming round, and on some loops the screen, as an AI coding round using HackerRank's built-in assistant. The transaction rule parser is the reported question. Candidates who did well formed their own plan before prompting, read the assistant's code closely and added their own tests.

Since January 2026 the Amazon OA often pairs an algorithm problem with about 60 minutes of debugging an existing repository. The assistant will point you at files but will not write the fix. Reproduce from the README and read the failing tests before changing code.

Anthropic runs a build-an-agent round, where you write the agent loop yourself, and a separate AI-assisted round where you use Claude Code to review and fix pull requests in an unfamiliar repository. A recent onsite report describes an AI round that starts from fixing a bug.

Write a client that submits a query to Snowflake's API and polls its status. The follow-up asks what to do when the result is very large. The candidate, who was rejected, says Datadog expects every follow-up answered and wants you to explain what the AI did.

The invitation names two rounds, AI Coding Agent for coding, algorithms and debugging, and AI Coding Agent usage best practices, and asks for an IDE connected to a coding agent, with VS Code recommended. The candidate who reported back got two ordinary LeetCode problems, where using AI for tests was optional. Standards seem to vary by team.

Recruiter description, no question reported yet: 60 minutes, a multi-part open-ended problem, and a working MVP you can run locally. It is graded on clarifying requirements, breaking the problem down, working with the AI, reviewing its code, testing, and handling changed requirements. One recruiter said every onsite is moving to this format.

Two reports: a command-line robot that moves and turns, with batch commands, obstacles and a question about the right input and output design; and a text editor with append and remove, with follow-ups on abstractions and output formats such as binary, Morse and HTML.

A 60-minute OA in a game simulation where you implement one method that decides whether a sprite should bounce. One candidate found the assistant unreliable. Another recognised it as a variant of a known flappy-bird task and wrote the roughly twenty lines by hand.

A September report says Google's L4 loop added an AI coding round and a solution design round, while other candidates still report the traditional format. An Apple candidate was told in April that every candidate would now get an AI-assisted round on CoderPad. A SoFi candidate interviewed through Karat in a browser editor whose assistant could not edit the code. ByteDance reports AI-enabled system design rounds: a file upload service, a monitoring dashboard and an agent quota system.