Run a four-to-five round loop that assumes AI use, totaling 4-5 hours of candidate time. Replace the algorithmic coding round with a 60-minute AI-assisted working session on a realistic problem. Add a code comprehension round on a real codebase. Score on five signals: orientation, AI use, verification, scope management, and communication — not on whether the code compiled. Companies still running the 2021 loop are losing offers to companies that aren't.
The interview loop most engineering organizations are running in 2026 was designed for a world that no longer exists.
In that world, asking a candidate to implement Dijkstra's algorithm in 45 minutes told you something useful. Watching them invert a binary tree on a whiteboard told you something useful. Six rounds of algorithmic puzzles, two days of onsites, a separate "system design" round — that loop, painful as it was, mapped roughly to what engineers actually did.
It doesn't anymore. Claude can write Dijkstra in 8 seconds. Cursor can invert a binary tree before the candidate has finished reading the prompt. The work that filled the 2021 loop is now the work that AI does without supervision. The candidates worth hiring know that. The candidates worth filtering out either don't, or pretend they don't.
The companies that have noticed have already rebuilt their loops. Canva announced in mid-2025 that AI tools are expected in all engineering interviews. Google now allows Gemini in coding rounds and added a "code comprehension" round on real existing code. DoorDash replaced its algorithmic rounds with a 60-minute AI-assisted working session where candidates join a realistic project, ramp up, and start contributing. Sierra runs an AI-native interview by default. The companies running 2021's loop in 2026 are losing offers to companies that aren't.
Here's the loop that works now, what each round should actually test, and how to score it.
The new 4-5 round loop
The right shape for an engineering interview loop in 2026 is shorter, denser, and built around watching how a candidate operates — not just what they produce. Four to five rounds, totaling 4-5 hours of candidate time end-to-end:
Recruiter screen 30 min
Calibration on level, comp band, location, and the candidate's actual job-search context. Skip the take-home at this stage in 2026 — take-homes get filtered through Claude before the recruiter reads them, so they tell you nothing.
Tests: level fit, motivation, basic alignmentAI-assisted working session 60 min
A realistic problem the team actually faces, scoped so pure prompting won't solve it. The candidate uses any AI tool they want (Cursor, Claude, Copilot — pick one and provide it). The interviewer observes how they orient, decide, prompt, verify, and ship. This replaces the algorithmic coding round entirely.
Tests: AI fluency, judgment under uncertainty, verification disciplineCode comprehension 60 min
A real (or realistic) codebase the candidate has never seen. They have 60 minutes to read it, find a bug, propose a fix, and articulate the trade-offs. AI tools available. This is the round that catches candidates who can prompt but cannot read code.
Tests: reading comprehension, debugging, architectural intuitionSystem design 45-60 min
Unchanged from 2021 in shape, but ask realistic questions, not whiteboard fantasies. "Design Twitter" tells you nothing. "We're seeing a 30% latency spike on our checkout flow at peak traffic — walk me through how you'd investigate and what you'd ship in the next sprint" tells you a lot. AI tools are usually not used in this round.
Tests: systems thinking, trade-off articulation, operational maturityCulture and values 45 min
Conducted by someone outside the hiring team. Discusses how the candidate operates — collaboration, disagreement, feedback, what they'd want from a manager. The right interviewer here is a Senior or Staff engineer who can detect when a candidate is reciting answers vs reflecting honestly.
Tests: values alignment, self-awareness, communicationThat's it. Five rounds. Most companies don't need more. If you're running 6, 7, or 8 rounds, you're testing the same dimensions repeatedly and your offer acceptance rate is suffering for it.
What each round actually tests (and what it doesn't)
The shift between 2021 and 2026 is that each round now tests a much more specific signal. The old loops blurred everything — "coding" rounds tested algorithms, syntax, debugging, communication, and stress tolerance all at once, with no clean way to score any of them. The new loops decompose the signals so each round produces clean evidence on one or two dimensions.
The AI-assisted working session is the most important round
This is the round that has replaced the algorithmic coding interview, and the one most companies still get wrong. The instinct is to give the candidate a contained problem — "build this small feature" — and watch them ship. That's not what this round is for. The whole point is that AI will ship the small feature in five minutes. The problem needs to be substantial enough that how the candidate works matters more than whether they finish.
The DoorDash framing is the cleanest: "join a realistic project, ramp up, and start contributing." The candidate is dropped into a problem with too much surface area to solve in 60 minutes. The interviewer is watching:
- Does the candidate read before they prompt? Or do they paste the prompt into Cursor and start firing?
- How do they decide what to delegate to AI and what to reason through themselves?
- When AI produces something, do they read it critically or accept it?
- Do they catch the bug AI introduced in the third generation?
- When they hit a wall, do they reset and rethink, or keep prompting harder?
- What do they cut when they realize they can't ship all of it?
The output of this round is not "candidate completed the task." The output is a paragraph from each interviewer describing exactly how the candidate operated. That paragraph is the signal.
Code comprehension catches the AI-native fakers
The dirty secret of the 2026 hiring market is that there's a cohort of candidates who can prompt their way through any greenfield problem but cannot read 200 lines of code they didn't write. They pass loops without this round and fail in week three on the job. The code comprehension round exists to catch them.
Give the candidate a real (or realistic) codebase — one or two services, 5-10K lines, with a bug that requires understanding the data flow to find. They have 60 minutes. AI tools available. The bug should be one that AI alone can't find without the candidate orienting first — usually a subtle interaction between two components, or a race condition, or an incorrect assumption in a comment that contradicts what the code does.
What you're watching: do they read the README? Do they trace the failing path? Do they ask AI to explain the architecture before asking it to find the bug? When AI produces a wrong answer, do they accept it?
Candidates who pass this round have a very specific skill: they can hold a foreign codebase in their head long enough to make a useful contribution. That's most of the job.
System design rewards specificity
The classic "design Twitter" or "design YouTube" prompt has aged badly. It rewards candidates who've memorized the answer and punishes candidates who reason from first principles. In 2026, replace these with operational problems your team actually faces.
"Our event pipeline is hitting 3x its design throughput and we're starting to drop events. Walk me through how you'd diagnose, what short-term mitigations you'd ship this week, and what long-term re-architecture you'd propose." That question is impossible to memorize and reveals exactly how the candidate thinks about scale, trade-offs, and operational maturity.
The scoring rubric: five signals, 1-4 scale
The scoring rubric for the new loop should be brutally simple. Five signals, four points each, every interviewer scores every signal they observed. No "overall vibe" score. The hire / no-hire decision falls out of the rubric.
| Signal | What "4" looks like |
|---|---|
| Orientation | Reads the problem and codebase before writing. Asks clarifying questions that show they're modeling the system. Forms a working hypothesis before reaching for tools. |
| AI use | Delegates well-defined subtasks to AI. Maintains control of the overall solution. Knows when to bypass AI entirely and reason directly. |
| Verification | Reads AI output critically. Tests it. Catches errors. Treats AI as a fast junior, not an oracle. |
| Scope management | Cuts ruthlessly under time pressure. Communicates trade-offs out loud. Ships the most important thing first. |
| Communication | Articulates what they did and why. Explains trade-offs to the interviewer without prompting. Self-corrects when challenged with evidence. |
A strong hire scores 3+ on all five. A weak hire scores 4 on one signal and 1-2 on the others. The dangerous candidate to hire is the one who scores 4 on AI use and 1 on verification — they ship a lot, fast, but they ship the wrong things.
Calibration check: Run your existing engineers through this rubric on a recent project. The Senior Engineers should score 3+ on all five signals. The Staff Engineers should score 4 on at least three. If they don't, recalibrate the rubric — either the bar is wrong, or the levels are mislabeled internally.
The three things to stop doing
Most of the work of redesigning an interview loop in 2026 is subtraction, not addition. Three things most companies are still doing that should stop:
Stop banning AI tools
Some companies still ban AI tools in interviews "to test what the candidate can do alone." This made sense in 2022. It doesn't now. The candidates you most want to hire use AI fluently in their actual work. Banning AI in the interview signals that you don't understand how the role is performed in 2026, which is the worst possible message to send to a top candidate. They'll accept an offer from Canva or Google or Sierra instead.
Stop the algorithmic coding round
Inverting a binary tree, reversing a linked list, implementing dynamic programming on graphs — this round used to differentiate engineers because the work was hard. AI has made it trivial. The candidates who still solve these problems fastest are the ones who've drilled LeetCode for six months. The candidates who become great engineers are not necessarily the same cohort. The round filters for the wrong skill.
Stop the 6+ round loop
Top engineering candidates in 2026 have multiple offers within two weeks of starting a search. Loops that take 4-5 weeks lose them to loops that take 1-2. If your loop has 7 rounds, half of them are testing the same signals. Cut to 4-5 and ship faster offers. Your acceptance rate will go up, your candidate experience score will go up, your time-to-hire will go down. These are not subtle effects.
The 2026 candidate-experience killer: A 4-week loop with a banned-AI policy. You will lose every candidate worth hiring to a competitor running a 2-week loop that assumes AI use. Brand doesn't compensate for this anymore — the loop itself is the brand signal.
What top companies are actually shipping
The specific tactical changes leading engineering organizations have made in 2025-2026:
- Canva — expects AI tools (Copilot, Cursor, Claude) in all engineering interviews. Tests "strategic use," not raw output.
- Google — allows Gemini in coding rounds. Added a code comprehension round on real codebases.
- DoorDash — replaced traditional algorithmic rounds with a 60-minute AI-assisted working session. Candidates join a realistic project and contribute.
- Sierra — runs an AI-native interview by default. The expected workflow is "join an AI agent project, ramp on the codebase, ship a small change."
- Anthropic — tests Claude fluency directly. Candidates are expected to use Claude in problem-solving rounds.
- Cursor — the entire interview happens in Cursor. The candidate's prompting patterns are themselves a signal.
The throughline across all of these: the interview increasingly looks like the job. Not a synthetic test of skills the engineer will never use again, but a 60-90 minute slice of what they'll actually do on Tuesday morning.
How to roll this out without breaking your pipeline
You don't have to rebuild the entire loop in one quarter. The minimum-viable migration most engineering organizations can run this quarter:
- Replace one algorithmic round with an AI-assisted working session. Pick the round that produces the least useful debrief notes today. Build one realistic 60-minute prompt. Train your interviewers on the new scoring rubric. Ship it.
- Add the code comprehension round. Use a sanitized version of your real codebase. Pick a bug you've already shipped a fix for. Calibrate against the engineer who originally found and fixed it.
- Cut one round. Look at your loop. There is a round where every candidate passes. That round is signal-free. Cut it.
- Update the rubric. Move every interviewer to the five-signal rubric. Run a calibration session with three recent hires to align on what 1/2/3/4 looks like for each signal.
This is roughly four to six weeks of work for a hiring team. The payoff — better hires, faster loops, higher offer acceptance — shows up within the first quarter.
And the meta-point: the candidates who are looking for jobs in 2026 read articles like this one. They know how the loop is supposed to work. When they sit down for your interview, they're scoring you as much as you're scoring them. The loop you run is the strongest signal a candidate has about how your engineering org actually operates. Make sure it's the signal you want to send.
Frequently Asked Questions
Reach engineers who can actually use AI tools
JobsByCulture lists your roles in front of engineers who research how teams hire before they apply. Show off your interview loop, your written engineering ladder, and your culture — the signals candidates use to filter in 2026.
Post Engineering Roles → What engineers check first →