Meta Onsite Interview Loop: What Each Round Is Really Testing
- The Meta onsite interview loop scales by level: E4/E5 get four rounds, E6 gets five with a second design round, E7+ skips standard coding entirely
- Coding round demands two problems in 45 minutes: finishing the first in 15-20 minutes is the only way to attempt the second
- The AI-assisted round scores tool judgment, not code output: catching AI hallucinations and prompting deliberately are the evaluated skills
- System design splits into two formats: distributed infrastructure for backend-heavy roles vs. product architecture for features and APIs
- The behavioral round is the only unrecoverable fail: a No Hire there cannot be offset by strong coding or design scores
- Meta's confidence score changes how incomplete rounds are weighted: narrating clearly and responding well to hints raises confidence even on imperfect solutions
You land the Meta recruiter call. You clear the phone screen. And suddenly you have four or five back-to-back interviews on the calendar with a company that processes hundreds of thousands of user actions per second and could, in theory, test you on all of them.
The Meta onsite is different from most loops you have practiced for. It is faster, more structured, and the rounds are not equal weight. One of them you genuinely cannot fail and recover from. Before you spend six weeks grinding LeetCode in a dark room eating cold pizza, understand what each round actually evaluates.
The Loop Scales With Your Level
Meta does not run the same gauntlet for every candidate. The structure scales with target level.
| Level | Standard Coding | AI-Assisted Coding | Design (System / Product) | Behavioral |
|---|---|---|---|---|
| E4 (Mid-level) | 1 | 1 | 2 | 1 |
| E5 (Senior) | 1 | 1 | 2 | 1 |
| E6 (Staff) | 1 | 1 | 2 (architecture + product) | 1 |
| E7+ / M1 | 0 | 1 | 2-3 | 1 |
Round counts and levels match the published breakdown in the interviewing.io Meta hiring guide. Meta rolled the AI-assisted coding round out in October 2025 and has been expanding it across SWE roles since, per Hello Interview's writeup. E7 and manager-level candidates skip the standard coding round entirely and get only the AI-assisted version. For everyone else, the standard round runs alongside the new one.
Two Problems, Forty-Five Minutes. Good Luck.
The standard coding round is 45 minutes. You are expected to solve two problems. Not back to back at a relaxed pace with snacks and a motivational playlist. You need to finish the first in 15 to 20 minutes before the second becomes viable.
The defining feature of Meta's coding round is speed. Google asks one problem per round and expects architectural depth. Meta asks two and expects execution. Finish the first in 30 minutes and you have zero time for the second, which reads as a weaker signal than finishing both partially.
Difficulty skews medium-to-hard at onsite, per the interviewing.io Meta guide, which notes that mediums dominate the loop and hards show up more often as you advance through the process. Common patterns: arrays, hash maps, trees, graphs, sliding window, and dynamic programming. Meta interviewers favor problems with a clear brute-force starting point so you can demonstrate your reasoning as you optimize toward something better.
You code in CoderPad. No IDE autocomplete. Write clean, readable code and narrate enough that the interviewer can follow without interrupting you. If you need help with the narration itself, our technical interview communication guide walks through what to say at each beat.
Two things Meta explicitly scores that LeetCode practice cannot train: testing and communication. Saying "let me verify this with a concrete example before I move on" earns points on Meta's rubric. It sits right next to time complexity analysis as a scored behavior. Dry-run your code on the input you defined at the start. Name at least one edge case. These are independent signals, not extras.
The pacing problem trips people up more than the problem difficulty. Practice the time split explicitly: three minutes to understand the problem and ask clarifying questions, two minutes to sketch a brute force, ten to twelve minutes coding the optimal solution, and three minutes verifying. That leaves a buffer. The buffer matters because Meta interviewers do interrupt, and recovering clean costs time.
The AI-Assisted Round Is Testing Something Sneaky
Meta started rolling this out in October 2025 and has expanded it across SWE roles since. The round is 60 minutes, longer than the standard because the tasks are more open-ended.
The environment is a three-panel CoderPad layout: a file explorer on the left showing a multi-file codebase, a code editor in the middle, and the AI chat plus problem statement on the right. The AI responds in its chat panel but cannot directly edit your files. Available models, per Hello Interview's writeup, include GPT-4o mini, GPT-5, Claude Haiku and Sonnet (3.5/4/4.5), Claude Opus 4, Gemini 2.5 Pro, and Llama 4 Maverick. You can switch between them mid-round.
The session moves through three progressive phases rather than three independent problem types:
- Bug fixing: a broken implementation you diagnose and repair to warm up.
- Core implementation: building the main feature on top of that fixed code.
- Optimization: speeding the implementation up under larger test cases.
The AI is there to help you move faster, not to solve the problem for you. Hello Interview's breakdown frames it bluntly: "read every diff, test it, and modify or reject it." Interviewers watch how you prompt the AI, evaluate its output, and catch its mistakes. A candidate who pastes AI suggestions without reading them fails even if the code compiles. The evaluation dimensions are the same four as the standard round: problem solving, code quality, verification, communication. The difference is the tool-assisted environment.
Every candidate who copies the AI's first suggestion without a second glance.
Prepare for this round by practicing with an AI coding assistant open alongside a problem. Prompt it deliberately, read what it returns, and find the errors. Getting fast at catching AI hallucinations and explaining why they are wrong is exactly the skill this round measures. For more context on how this format is evolving across the industry, see AI-Enabled Coding Interviews: What's Actually Changing in 2026.
System Design: Know Which Version You Are Getting
E5 candidates get one design round. E6 gets two. The format depends on your role and team, and the distinction matters.
The System Design version focuses on distributed systems at Meta's scale: feed ranking infrastructure, notification pipelines, messaging architecture, real-time data systems. You are expected to reason about fan-out strategies, caching layers, consistency tradeoffs, and what happens at billions of users. Interviewers here have built production systems handling hundreds of thousands of writes per second. Vague handwaving about "using Kafka" will not satisfy them.
The Product Architecture version focuses on how a product evolves as requirements change. API design, client-server interaction, schema decisions, maintainability. Think "design the data model and API for a marketplace listing feature" rather than "design the Kafka pipeline for 500K writes per second." The bar is still high, but the lens is different.
Both formats run 45 minutes and score on four dimensions: problem navigation, solution design, technical excellence, and communication. The most common mistake in Meta's design rounds is skipping the requirements conversation and jumping straight to boxes and arrows. Use five minutes to establish scope, rough user scale, and the hardest constraints. The clarifying-questions guide covers the exact questions worth asking up front. A design dimensioned for the wrong scale is wrong regardless of how elegant the architecture is.
See System Design Interview: What to Expect, How It's Scored, and How to Stand Out for a longer treatment of how to run the 45-minute clock.
The Behavioral Round: The Only One You Cannot Come Back From
Every round matters, but behavioral is different in a specific way. A weak coding round can be discounted if the committee believes the question was off-pattern or unusually hard. A clear "no hire" out of behavioral is much harder to argue around. The interviewer's call is not literally final (debrief committee owns the decision, per interviewing.io) but in practice a bad behavioral packet sinks the loop.
Per interviewing.io's evaluation breakdown, Meta scores behavioral against eight focus areas, not the vaguer "tell me about a time" prompts a candidate hears:
- Motivation: why you work on what you work on, and what pulls you toward impactful problems.
- Proactivity: going after hard, unowned problems before someone tells you to.
- Working in unstructured environments: taking ownership when scope is fuzzy and no one wrote the spec.
- Perseverance: pushing through real blockers without quietly dropping the work.
- Conflict resolution: disagreeing with peers or leadership, reaching alignment, listening before persuading.
- Empathy: reading what other people actually care about, including users.
- Growth: self-awareness about strengths and weaknesses, and evidence the gaps shrink.
- Communication: clean storytelling under follow-up pressure, including this interview itself.
The round is 45 minutes. Your interviewer will ask three or four questions and probe each one deeply. Pick stories with multiple layers so you survive follow-up: "What would you have done differently?" "Who disagreed with you?" "How did you know it was working?" Stories that collapse under the first follow-up are not ready. Go back to the drawing board.
Use STAR format, but compress Situation and Task to one or two sentences. Meta interviewers are tracking eight distinct signals in 45 minutes. They do not need three minutes of context before your action begins. Get to what you actually did, fast.
A story is ready when you can answer "what would you have done differently" without undermining the original decision. If walking back your action destroys the story, find a different story. The follow-up is not a trap. It is how interviewers distinguish candidates who learned from experience from candidates who got lucky.
The stories that score highest share a through-line: you took initiative, moved quickly with incomplete information, and communicated directly when it was uncomfortable. Those are the behaviors Meta's values describe whether you cite them by name or not. A strong behavioral answer does not reference the company's values. It demonstrates them.
For a full breakdown of how behavioral rounds are scored and what the rubric actually rewards, see The Software Engineer Behavioral Interview Isn't a Culture Fit Screen.
The Confidence Score: A Meta-Layer Nobody Prepares For
Meta is the only major tech company that asks interviewers to attach a confidence level alongside their hire or no-hire recommendation, per the interviewing.io Meta guide. If an interviewer gives you a hire rating with low confidence because the problem was unusually hard or the session dynamic was off, that round carries less weight when the hiring committee reviews the packet. The same works in reverse: a no-hire with low confidence is easier to overturn.
One rough round is survivable if your other rounds are strong. Candidates who narrate clearly, respond well to hints, and stay composed under pressure generate higher-confidence recommendations even when their solutions are incomplete. Composure is a trained behavior. It comes from repeated practice under realistic conditions, not from telling yourself to relax in the moment.
Running timed sessions on SpaceComplexity to practice explaining your thinking out loud builds exactly the kind of muscle that produces high-confidence signals in the real round.
Stop Preparing for One Round When There Are Four
Meta's loop weighs coding, AI-assisted coding, design, and behavioral as roughly equal, so the weakest of the four sets your ceiling. Match your prep accordingly, weighted by what is hardest to fix quickly.
Coding: Solve LeetCode medium problems in under 20 minutes. Once that feels consistent, do two back to back on a single timer. Add an explicit verification step every time, out loud, before you declare done. For the AI-assisted round, run at least one session per week with an AI assistant open alongside a problem. Prompt it, read the output, find the bugs.
System design: Run three or four full 45-minute mock designs at Meta's scale. Internalize the patterns for feed systems, notification infrastructure, and messaging. Practice scoping the problem out loud before drawing anything. If you are targeting E6, prepare two different architectural angles for each problem, not just one polished answer.
Behavioral: Write four or five strong stories that collectively cover the eight focus areas. Read each one aloud and add a follow-up question after every sentence in the action section. If the story falls apart under one follow-up, find a different story. Practice spoken, not written. The version that sounds good in your head and the version that holds up out loud are not the same thing.