Meta AI-Native Coding Interview: 3 Experiences

Shared by janebeingjane ·
Company
Meta
Reports
3 interview reports
Level
Full-time
Stages
Onsite

Meta's 2026 AI-native coding onsite had two rounds where engineers solved hard algorithm problems with an AI model, with correctness checked before speed.

At a glance

  • Process: an ordinary coding phone screen with no AI, then two AI-native coding rounds at the onsite beside two system design rounds and behavioral (01, 02).
  • Asked recently: hard algorithm problems inside mostly written code, with test suites including a stress test (02); a card-game round asked for new benchmark tests over 1,000+ games (02).
  • Difficulty: 01 failed a round by optimizing AI code that still had a bug; 02 passed by checking correctness first. The MLE was waitlisted (03).
  • Tools: Claude and GPT models to choose from; in 01 every prompt needed the interviewer's approval first.

Coverage: Experienced full-time hires who were changing jobs: two software engineers (one at E6, one whose level wasn't given) and one machine learning engineer. All three describe the onsite, and the two software engineers also mention the phone screen. The interviews took place in July–September 2026 and October–December 2026, and the reports went up between September 15 and November 10, 2026. No locations were given.

Roles in this article

  • Software engineer, 2 reports: Kitchen Orders, Maximize Unique Characters, friend recommendations and a 36-card game in the AI-native rounds, plus four system design questions.
  • Machine learning engineer, 1 report: two AI coding questions named without detail, plus two ML system design rounds.

Software engineer: Meta AI-native coding onsite (2 reports)

What came up

Both software engineers had an ordinary coding phone screen and then two AI-native coding rounds at the onsite, with Claude and GPT models to choose from. Both accounts keep returning to one tension: getting the AI's code right before trying to make it fast.

Topic Reports Last seen
Two AI-native coding rounds at the onsite, Claude or GPT models 2 of 2 Nov 2026
Checking the AI's code for correctness before optimizing 2 of 2 Nov 2026
Classic system design in the other onsite rounds 2 of 2 Nov 2026
Coding phone screen with no AI 2 of 2 Nov 2026
Scheduling steps across workers to finish fastest 1 of 2 Nov 2026
Picking words with no shared characters 1 of 2 Nov 2026
Friend recommendations on a large graph 1 of 2 Sep 2026
Card-game search with a benchmark over many games 1 of 2 Sep 2026

Each report can count toward several rows. Both reports also had a behavioral round, with no questions named.

Questions we found

AI Coding

Question Reports Last seen
Kitchen Orders: assign each recipe step to one of three chefs so the whole order finishes as early as possible 1: 01 Oct–Dec 2026
Maximize Unique Characters: choose words that share no characters so they cover as much of an alphabet as possible 1: 01 Oct–Dec 2026
Mutual-Friend Recommendations: recommend friends from users and friend pairs, tested up to about 100k users 1: 02 Jul–Sep 2026
Partition a Card Deck into Fifteens: a 36-card game of picking three cards that sum to 15 1: 02 Jul–Sep 2026

System Design

Question Reports Last seen
Ads click aggregator (Design an Ad Event Measurement Platform) 1: 01 Oct–Dec 2026
News feed (Design a Following News Feed Service) 1: 01 Oct–Dec 2026
Game score leaderboard that shows the nearest scores after a game (Design a Social Game Score Leaderboard) 1: 02 Jul–Sep 2026
Ride sharing (Design Ride-Hailing Dispatch and Trip Lifecycle) 1: 02 Jul–Sep 2026

The last two were Meta's "AI-enabled SWE system design" rounds.

2 detailed candidate timelines

01: E6 engineer, failed after missing a bug in the AI's code

Report date: Nov 10, 2026 · Interview period: Oct–Dec 2026 · Role: Software engineer, E6 · Experience: changing jobs, referred · Result: Rejected

Phone screen

  1. Traditional coding with no AI. The problems weren't named.

Onsite: AI coding round 1, Kitchen Orders

  1. A kitchen has to cook one order made of several dishes. Each dish follows a recipe, which is a fixed sequence of station steps.
  2. Every step goes to one of three chefs, and the goal is to finish the whole order as fast as possible.
  3. Sending each step to its fastest chef is only a start. The order in which the steps are scheduled matters too: in the round's three-dish example, a better order saved 30 seconds.

Onsite: AI coding round 2, Maximize Unique Characters

  1. The input is a list of words and an alphabet. The task is to pick a set of words that covers as many characters as possible, with one rule: no two picked words may share even one character.
  2. This round went wrong. The candidate didn't read the AI's code closely, missed a bug, and moved straight on to optimizing it. With about ten minutes left, the interviewer tested correctness and found the problem. The candidate fixed it at the last second, but still failed.

How the AI rounds worked

  1. Both Claude and GPT models were available.
  2. The candidate didn't have to write their own tests.
  3. The models felt like weaker versions of the public ones, and what they wrote wasn't always free of bugs.
  4. Before sending a prompt, the candidate had to explain the plan and get the interviewer's approval.

Onsite: system design and behavioral

  1. System design 1 was an ads click aggregator.
  2. System design 2 was a news feed.
  3. The behavioral round used the usual questions; none were named.

The candidate said every question in this loop had been reported by other candidates before.


02: Engineer who passed, with a detailed playbook for the AI rounds

Report date: Sep 15, 2026 · Interview period: Jul–Sep 2026 · Role: Software engineer, level not given · Experience: changing jobs, referred · Result: Passed

Phone screen

  1. Two LeetCode-style problems, one easy and one medium, with no AI.

Onsite format

  1. Two "AI-native coding" rounds, two "AI-enabled SWE system design" rounds and one behavioral round. Those were Meta's own names for the rounds, according to the candidate.
  2. The coding rounds were harder algorithm problems handed over as a whole folder of mostly written code. The candidate, working with the AI, filled in the core logic.
  3. Each round was a fresh CoderPad session of about 40 minutes. The candidate worked in Java with Maven.
  4. The code ran in a shared container with roughly 1.5 CPU cores and 2 GB of memory, so multithreading or several agents at once weren't an option. The tools were sometimes flaky: one could report a timeout even though the edit had gone through.
  5. The models offered were Claude 4.8 and GPT 5.5. GPT 5.6 was already out, but the interview only had 5.5, and the candidate guessed the interview models trail public releases by a month or two.

Onsite: AI coding round 1, friend recommendations

  1. Given users and pairs of friends, implement a friend-recommendation strategy: friends of friends, popular users, and similar signals. The closest practice version is Mutual-Friend Recommendations, which keeps the mutual-friend ranking core.
  2. There were three test suites. The third had about 100k users and 1.5 million friend pairs, so it worked as a stress test.

Onsite: AI coding round 2, the 36-card game

  1. A deck of 36 cards with values 1–9. A move picks three cards that sum to 15.
  2. Only about 16 cards sit on the table at a time, and three new ones are dealt after each pick. You win if all 36 cards get used, which means 12 groups.
  3. The candidate had to write new tests that benchmark the algorithm over a large number of games, more than 1,000.
  4. Their final result: a 99.97% win rate on the 10k-game case, in 40 seconds.
  5. Partition a Card Deck into Fifteens practices the core question of whether a fully visible deck can be split into fifteens. It doesn't simulate the 16-card table or the benchmark.

How the candidate ran the AI coding rounds

  1. They opened each session with a prompt that told the AI about the time limit and the environment, and re-entered it every round since each one started fresh.
  2. They had the AI read the code and the task first, then restate the goal and what counted as passing.
  3. Correctness came first, then speed. They aimed straight for the final optimized algorithm instead of writing a brute-force version first.
  4. They worked in chunks of 2–5 minutes covering plan, implement and test. Any AI reply that would take more than about three minutes of work had to be broken down first, with the expected gain, effort and risk stated before any optimization.
  5. They kept outputs short and skipped test runs that weren't needed, since every Java run meant a build. No background or long runs, always a timeout, single-threaded only.
  6. They re-read files after every edit, because of the flaky timeout reports.
  7. For speed work they used a small measurement harness with a fixed seed and sample sizes of 20, then 100, then the full set. No claim of a speedup without a before-and-after on the same sample, and a running changelog.
  8. They warned that newer models may start the big test suites on their own and burn time, or hang before any optimization happens.

Onsite: AI-enabled system design

  1. The AI here was just a chat window offered "to unblock yourself", and the candidate found it close to useless. Neither interviewer gave it a real use.
  2. Design 1 was a video game score leaderboard: players, games, friends and high scores, and after a game, show the scores closest to the player's. The candidate saw it as a range search over a large dataset.
  3. Design 2 was ride sharing.

Onsite: behavioral

  1. One round. No questions were named.

Machine learning engineer: Meta AI coding and ML design onsite (1 report)

What came up

The one MLE report lists two AI coding questions that the author called frequently reported, without describing either, and two ML system design rounds: weapon-sale detection and video copyright matching.

Questions we found

AI Coding

Question Reports Last seen
Ordering system for a menu (named only) 1: 03 Jul–Sep 2026
"Unique string" problem (named only) 1: 03 Jul–Sep 2026

ML System Design

Question Reports Last seen
Design Video Copyright Matching Against a Reference Library 1: 03 Jul–Sep 2026
Design Weapon-Sale Ad Detection and Review 1: 03 Jul–Sep 2026

1 detailed candidate timeline

03: MLE onsite, waitlisted

Report date: Nov 6, 2026 · Interview period: Jul–Sep 2026 · Role: Machine learning engineer · Experience: changing jobs, came through a headhunter · Result: Waitlisted

Onsite: AI coding

  1. Two questions, both of which the author called frequently reported: an ordering system for a menu, and a "unique string" problem. Nothing else about them was shared.
  2. The names sound close to Kitchen Orders and Maximize Unique Characters in report 01, but this report gives too little to say they are the same tasks.

Onsite: ML system design 1, weapons detection

  1. The author called it a heavily repeated question and didn't remember anything unusual. Practice version: Design Weapon-Sale Ad Detection and Review.
  2. The interviewer was friendly the whole way, barely pushed back on the design, and now and then mixed in questions on ML basics.

Onsite: ML system design 2, copyright detection

  1. How do you check whether an uploaded video duplicates one in a copyright library? Practice version: Design Video Copyright Matching Against a Reference Library.
  2. The main focus was comparing videos of very different lengths. The author had no experience with this and improvised.

Patterns across the reports

These patterns come from the two software engineer reports (01 and 02). Report 03 is a different role, so it stays out of the comparison.

  1. AI only at the onsite. Both had a phone screen of ordinary coding with no AI, then two AI-native coding rounds at the onsite next to two system design rounds and a behavioral round.

  2. Tests were mostly provided, with an exception. Report 01 didn't have to write tests. In report 02, the friend-recommendation round came with three test suites, the last one a stress test, while the card-game round asked the candidate to write new benchmark tests over more than 1,000 games.

  3. Correctness before speed. Report 01 lost a round by optimizing AI-written code that still had a bug the interviewer then found. Report 02 checked correctness first and only claimed speedups measured on a fixed sample.

  4. The interviewer stays in the loop. In report 01, every prompt needed the interviewer's approval of the plan first. Report 02 described the interviewer watching the whole exchange with the AI, which made a wrong turn costly.

  5. System design stayed classic. Neither report describes AI changing the design questions. In report 02, the AI on offer in those rounds was a chat window the candidate barely used.

Common questions

What is the Meta AI-native coding interview?

Two onsite rounds where you finish hard algorithm problems with an AI model, then pass provided tests and stress tests (01, 02).

How do you pass the Meta AI coding round?

Candidate 02, who passed, checked correctness before optimizing; 01 failed after the interviewer found a bug in AI-written code.

Keep practicing

AI coding rounds at other companies

More companies with AI rounds on FastPrep:

Across companies: 2026 AI Coding OAs at 15 Companies: 32 Reported Problems and 2025–2026 AI Coding Phone Screens and Onsites at 43 Companies: 76 Reported Problems.

Sources: Anywhere useful interview info can be found.

0 total reactions. Your reactions: none.

Comments

0

No comments yet — be the first to share what you know.