What candidates say the round looked like, summarised in our words, each linked to where it came from. Meta hasn’t confirmed these details unless the source says “Official page”.
Background on the format: our Meta guide.
Had this round? Add your report →
The candidate passed all 4 test cases in the AI coding round but was still asked to do a follow-up interview. The interviewer explained that the AI coding round had gone well and the weak spot was the classic coding round, but the follow-up would use the AI coding format anyway, because all roles were now ‘AI native’.
For an ‘AI native’ software engineer role at E5 or E6, the recruiter described the loop as 2 AI-enabled coding rounds, 2 system design rounds and a behavioural. A Meta engineer in the thread said AI-native coding means a bigger problem where you may use AI, simulating day-to-day work, and that using it is your choice. Design and behavioural rounds usually decide the level.
E4 Android loop that ended in rejection. The AI round was the maze solver task reported by others: the main algorithm was a BFS, and the work was getting the failing tests to pass. The rest of the loop was two classic coding problems, behavioural questions, and a mobile system design prompt with no mocks, where the candidate had to work out the requirements themselves.
Watch the name: for a research engineer role, Meta’s ‘AI coding’ round is not the AI-enabled round. This candidate got two questions with no AI assistant: one very common LeetCode problem, and implementing a simple 2D convolution from scratch. Ask your recruiter which of the two formats you have.
Strong hire on every round, including AI coding. The first two parts went quickly, solved verbally or with a dry run. On the next one the AI produced wrong code, and fixing it took most of the remaining interview, but they got it working. There was no time left for the third, so they only dry ran and explained it.
ML research scientist loop. The coding with AI round gave a full project of 4 or 5 files with tests and an AI agent, with bugs spread across files to fix until the tests pass. The candidate passed 11 of 12 tests before the interviewer said there was no need for the last one. Twice when they tried to query the AI, the interviewer stopped them and asked them to look closer themselves, so in the end they didn’t use the AI at all. They did well here; the rejection came from other rounds.
A new grad preparing for the loop noticed that Meta’s CoderPad practice question had added more models to the AI Assist list: GPT-4o mini, GPT-5, Claude Haiku 3.5 and 4.5, Claude Sonnet 4 and 4.5, Gemini 2.5 Pro and Llama 4 Maverick. Earlier candidates had only small models. Check the list in the practice pad before your round.
Successful Software Engineer, Machine Learning E4 loop. The AI-assisted round was a mini project with 4 tasks of rising difficulty. The hardest part was understanding the codebase at the start: the first task takes longest because you’re learning their helper functions. The interviewer asked them not to use AI on the first task, and they ended up coding by hand throughout and finished 3 of 4. Tip: prioritise passing tests over clean code, and explain how you’d refactor.
E5 infra loop in the US. The AI coding round, then still a pilot, was a speed run through 4 incremental stages. You aren’t expected to write a full algorithm from scratch, but you need common techniques, such as handling conditions during a traversal, at your fingertips. Don’t hand the problem solving to the AI: define the requirements yourself, then ask it for code. The candidate’s stages were about edge cases in traversals, they liked the round, and they finished all 4 in time.
Same shape as other reports: first fix two failing test cases, then implement a solver function. The project had four files and the candidate wrote code in three. The hard file needed a search over a long word list; they wrote a backtracking solution, explained its complexity, tried a bitwise optimisation when asked for better, and ran out of time, so they described further optimisations out loud. The fourth file barely mattered.
E4 product loop: phone screen hire, coding hire, system design strong hire, behavioural hire, but AI coding was a no hire after solving only two levels. Meta scheduled a follow-up coding round two weeks later, where the candidate solved both questions but felt the interviewer didn’t follow their thinking. They didn’t get the offer.
Feedback on this candidate’s first AI round: the interviewer had been hinting, and the candidate hadn’t used AI on level 1. In the follow-up AI round, they were told not to use AI on level 1. Their point: without extra time to read through the files, nobody can know every line of a large codebase in an hour.
E4 product loop in the US. Behavioural, coding and architecture went very well, but the AI-enabled round had about 5 levels and the candidate finished 2 before getting stuck on the third. The AI was poor, the codebase was large, they went in without knowing the format, and they panicked when things stopped working. The hiring committee asked for a follow-up round because the rest of the feedback was strong.
SDE2 loop. The AI-enabled round was a string processing problem: first a debugging portion, then implementing a function that removes words from a list when they share letters. Time ran out right after that, so the candidate doesn’t know if more follow-ups existed. The interviewer was very friendly and helpful.
AI-enabled round built around a card game: draw N cards into a grid and pick 3 whose values sum to 15. First, fix a failing test, much like the demo on Meta’s site; the bug was in validating invalid moves. Then implement a strategy that keeps picking valid triples until no cards remain. The AI was available throughout but was a small, not very reliable model; the choices were a small Llama, Claude Haiku 3.5 and one more. The candidate thinks the point is to see how much you verify, and would rather have done LeetCode. A Meta employee in the thread said only 4 vetted questions existed.
A four-part debugging task around solving a maze: fix a bug in a function that prints a valid path, fix a BFS implementation, then two parts that scale the problem up and need optimisation. Despite the round’s name, the interviewer asked the candidate not to use AI for part 1. When they did use it on the BFS fix, the model (they warn against Llama) gave wrong answers. A mismatch between the test output and the debug output cost extra time. Their take: it is really a debugging round, and AI is best kept as a last resort.
Product SWE onsite with one AI coding round alongside a classic coding round, system design and behavioural. CoderPad had an AI Assist tab limited to lightweight models (o4-mini, Claude Haiku, Llama 4 Maverick); the candidate picked Claude Haiku. The codebase had main.py, test.py, solve.py, utils.py and a data folder of word lists. Ticket 1: fix 2 of 4 failing util tests, minor logic errors. Ticket 2: implement solve.py to find a subset of words with a given property. After clarifying requirements, a careful prompt got working code in one or two tries. The catch came next: the solution was exponential, the interviewer asked for better, and the problem turned out to be NP-complete. Their advice: Meta grades the usual skills, not prompting, so show you use the AI responsibly.
E4 product role in Bangalore, in the first weeks of the new round. A codebase with 4 stages, each with failing unit tests to fix, graded on debugging, communication, code quality and navigating a codebase. The problem was the maze solver: walls, portals, and serialising the graph objects into a maze. Understanding the problem took 10 to 15 minutes, and part 1 was done 30 minutes in, with hints. The two BFS parts took another 20 to 25 minutes, ending at 3 of 4. The interviewer said solving everything wasn’t the point. The loop ended in a rejection.
E4 infra loop scheduled just as the format changed on 1 October 2025. The loop became one 45-minute coding round, one 1-hour AI-enabled round, one 45-minute system or product design round and a 45-minute behavioural. The recruiter knew little about the new round, and a Meta employee said only that you’re expected to solve the problem the way you would at work with AI. The candidate cleared the loop and moved to team matching.
Where it started: 404 Media reported internal Meta messages saying some coding candidates would be allowed to use an AI assistant during interviews, calling it ‘more representative of the developer environment that our future employees will work in’ and a way to make LLM-based cheating less effective. Meta also asked employees to volunteer for mock AI-enabled interviews. Business Insider reported that Meta confirmed the change, and candidates describe the round going live from October 2025.