Company guide · Meta

Meta wrote the new playbook.
The rest of the industry is following.

What changed in October 2025, why it matters for the next decade of engineering interviews, and how to prepare for the 60-minute three-phase project that replaced one of Meta's onsite coding rounds.

Last updated April 29, 2026. By the Nyrion team.

Why Meta did this first

Most of the industry has been agonizing for a decade over the same question: are we still going to ban AI in the room while every engineer on the team uses it daily? Meta's answer in October 2025 was: no. They put a real assistant in the interview environment, raised the bar on what candidates are expected to produce, and shipped.

Meta's internal announcement to mock candidates put the motivation plainly:

"Meta is developing a new type of coding interview in which candidates have access to an AI assistant. This is more representative of the developer environment that our future employees will work in, and also makes LLM-based cheating less effective."— Internal message to mock candidates, October 2025

Two motivations, both honest. First, alignment with reality: their own engineers ship with Cursor, Claude, and Copilot every day, so testing them in an artificial no-AI environment had stopped predicting on-the-job performance. Second, and quietly important, this format is harder to game with interview-prep cheating tools. When the interview expectsyou to use AI, hiding a second model in your other tab adds nothing.

Meta engineers talking publicly about the rollout were more philosophical:

"AI has fundamentally changed the way software is developed… Our modernized AI-enabled interview is a better representation of our work and of our mission."— Nicholas O., SWE, Programming Languages and Runtimes
"This new format enables engineers to actually execute their code, dive deeper into real practical problems, and leverage AI to brainstorm and iterate."— Peter C., SWE, Threads Product

Our take: Meta deserves credit for moving first on something that the entire industry knew was overdue. Shopify, LinkedIn, Canva, and Rippling have all started experimenting with similar formats in the months since. The conversation has shifted from "should we?" to "how soon?" — and that shift is downstream of Meta's October 2025 launch.

What the interview actually looks like

The most important structural change isn't "you can use AI." It's that the round is no longer two isolated algorithm problems. It's a single multi-stage project. You spend the whole hour inside one codebase, building one thing, with the requirements escalating as you go.

The three phases, in order:

Phase 1 — Bug fix in unfamiliar code

You're dropped into a small codebase you've never seen and asked to find and fix a bug. Reported examples include type casts gone wrong, off-by-one boundary errors, and incorrect conditional logic. The point isn't difficulty of the bug — it's whether you can navigate someone else's code under time pressure. This phase tests code reading, a skill LeetCode never measured.

Phase 2 — Core implementation

You build the main feature or algorithm. Candidates report ~120 lines of expected output, which is roughly 3–4× a typical LeetCode solution. You can prompt the AI for boilerplate, ask it to suggest approaches, paste error messages back, ask it to read sibling files. What you cannot do is hand off your judgment — Meta is grading whether you understand what you're shipping.

Phase 3 — Optimization under stress

Bigger inputs, edge cases, performance. The interviewer introduces new constraints — bigger N, concurrent calls, memory limits. This is where candidates most commonly run out of time. One reported finishing all three phases in 40 minutes; that's the upper bound, not the median.

What Meta is actually evaluating

Four competencies, all real, none about prompt engineering:

  • Problem solving. Picking the right algorithm, reasoning about edge cases, recognizing when a fix isn't actually fixing the problem. The AI accelerates execution but doesn't pick the approach for you.
  • Code quality. Clean, maintainable code that you demonstrably understand. If the assistant generated it and you can't explain why each line is there, that's the failure mode interviewers are watching for.
  • Verification. The single most predictive signal, based on candidate reports. Did you test before moving on? Did you read the AI's output instead of trusting it? Did you catch its mistakes? Meta's own evaluation rubric (per leaked guidance) says: "Test before using. Don't prompt your way out of it."
  • Communication. You're narrating a real engineering process. What you tried, what didn't work, why you chose this approach over that one. The interviewer needs to follow your reasoning — not your AI's.

The contrarian read

Half the candidate-prep internet is panicking that the bar dropped because the AI is now in the room. It didn't drop — it moved up. Three reasons:

  • You're now expected to read code, not just write it. That's a harder skill, not an easier one.
  • The implementation target tripled from one isolated function to a multi-file, ~120-line build under the same 60-minute clock.
  • You're being graded on how well you verify AI output. The candidates who fail aren't the ones who can't prompt — they're the ones who paste in code without reading it.

One candidate's quote captures it bluntly:

"If you don't know how to solve that problem by yourself, it's pretty difficult to use AI to solve that problem."— Anonymous E5 candidate, post-onsite report

Where this goes next

Meta is the first; every other large company will follow. That part is obvious. Less obvious is what the format will evolve into over the next 24 months. Our prediction — and we'll say so without hedging because we believe it — comes in three steps:

The bar keeps rising. The 120-line implementation target is calibrated for current-generation AI assistants. As the assistants get faster and more capable, the implementation target gets longer. The next round of this format won't ask for one feature in a small repo. It'll ask for several features across a medium one.

Problems shift toward production-shape. Today's phase-three optimizations are recognizable as algorithm problems with bigger inputs. Tomorrow's will look like real engineering: streaming aggregations, cache eviction strategies, schema migrations under load. The codebases will look like small services, not tutorial repos.

Eventually the interview question itself changes shape.We expect to see, within two to three years, prompts of the form "implement a recommender system that operates on terabytes of input data" — with a real (small) dataset, real tests, and a real AI assistant — replacing what we today call a system design round. The line between "coding" and "system design" interviews collapses, because once the assistant is doing the typing, the bottleneck stops being keyboard speed and becomes architectural judgment under time pressure.

Meta proved this format works. The next interesting question is which company pushes it furthest first.

How to prepare for it

Most prep advice for AI-enabled interviews is wrong because it focuses on prompt-engineering tricks. The actual high-leverage practice loop is:

  • Practice on real codebases, not snippets. The first phase is reading unfamiliar code. If your prep is all single-function LeetCode, you're optimizing for a skill that doesn't appear in this format.
  • Practice directing an assistant, not chatting with one. Get fluent at giving the model concrete tasks ("read parser.py, find where the buffer is reset inside the loop, propose a fix that preserves split records") instead of open-ended prompts.
  • Practice the verification reflex. Whenever the assistant proposes code, your default move should be: read it, identify what it's claiming, run a test that would distinguish that claim from a wrong one. Build the habit so it survives the time pressure.
  • Solo-solvable first, AI-accelerated second. Don't practice problems you couldn't solve unaided. The format punishes that skill profile — and so does Meta's "test before using" rubric.

Nyrion's catalog is built around exactly this loop. Each problem is a multi-stage repo, the agent's posture is execution-only (zero novel ideas), and tests gate progression. Filter by Meta on the catalog page to see problems specifically tagged for this format.

Frequently asked, briefly answered

When does the AI-enabled interview apply to me?

Currently invite-only at E5–E7 and M2. Junior roles still get the classic LeetCode loop. Expansion to broader levels is expected through 2026 — Meta's pilot framing strongly implies they'll roll it out further once metrics stabilize.

Will I still get a no-AI coding round?

Yes. The pilot replaces one of two onsite coding rounds. The other stays a classic no-AI LeetCode-style problem. So you still need classical algorithm prep — just less of it.

Which AI model should I use?

Whichever you've practiced with. The four supported models are intentionally limited (no internet access, basic chat only). The model isn't the bottleneck — your ability to direct it is. Pick one, get fluent, don't switch on interview day.

Do I have to use the AI?

No, but ignoring it penalizes you. The 120-line target in 60 minutes was calibrated assuming you use the assistant. Refusing to use it isn't a flex — it's leaving capacity on the table.

Is this easier than traditional LeetCode interviews?

No, and assuming it is, is the most common preparation mistake. The bar moved up. You're producing more code, in an unfamiliar codebase, against a tighter clock, while also verifying AI output. Meta knows the assistant exists — they calibrated for it.

Practice problems tagged for Meta

We curate problems that mirror this format — multi-stage, real codebase, AI agent in the editor, tests gating progression. Browse the catalog and filter by Meta to see them. No login required to look around; sign in when you want to actually solve one.

Sources