The traditional coding test, a timed algorithm puzzle in a blank editor, is losing ground. Most teams replacing it use task-based technical reviews, code walkthroughs, or live sessions where candidates work with their normal tools, including AI assistants.
Why the algorithm test stopped working
The classic whiteboard or HackerRank format was always a proxy. It measured a specific kind of performance, solving constrained puzzles under artificial pressure, not the broader judgement required in a real codebase. That proxy was already imperfect before AI tools became common. Now it is harder to justify.
A candidate who has memorised dynamic-programming patterns is not necessarily a better engineer than one who has not. And a candidate who can write a binary search from memory tells you nothing about whether they can read inherited code, catch a subtle model-generated bug, or make sensible decisions on an incomplete ticket.
The arrival of capable AI coding assistants widened the gap further. Engineers who rely on those tools daily, and use them well, may be slower at unassisted puzzle-solving while being faster and more reliable in practice. Testing the former penalises the latter.
What teams are using instead
There is no single replacement. Common approaches, often combined:
- Take-home task on a real codebase. The candidate receives a small, realistic problem, a failing test, a feature request, a refactor, and returns working code with a short explanation. The review focuses on structure and decisions, not just whether it runs.
- Code walkthrough. The candidate brings a recent piece of their own work and talks through it. This surfaces how they think about design, trade-offs and maintenance, things no algorithm test reaches.
- Live session with AI tools permitted. The interviewer watches how the candidate uses an assistant: what they prompt, what they accept, what they question. This is more informative than watching someone type unaided.
- Structured technical conversation. Instead of writing code, the candidate discusses a system design or debugging scenario. Useful for senior roles where judgement matters more than syntax recall.
Each has weaknesses. Take-homes cost the candidate time and can be outsourced. Walkthroughs depend on the candidate having shareable work. Live sessions require a skilled interviewer who knows what good AI use looks like.
What the shift reveals about hiring criteria
Replacing the test is easier than agreeing on what you are actually looking for. Teams that have moved away from algorithm tests often discover they had no clear framework for what the replacement should measure.
For roles where AI tools are central, the relevant skills include: knowing when to trust generated code and when to check it, reading unfamiliar code quickly, catching plausible-looking errors, and making decisions when requirements are thin. None of those appear on a LeetCode leaderboard. See how to assess an engineer's AI ability for a more detailed breakdown.
The related question of what reviewers should look for in AI-generated output is covered in how to review AI-generated code.
The risk of replacing one proxy with another
Some teams have moved to purely conversational interviews or portfolio reviews without adding any technical verification. That trades one problem for another. A candidate can discuss trade-offs fluently without being able to implement them, just as an algorithm test can pass someone who cannot build production software.
The better replacements combine a practical element, something the candidate actually produces or works through, with enough structure that different candidates are compared on the same dimensions. Without structure, interviews tend to favour confidence and communication style over technical competence.
What we test for
Miyagami's vetting does not use algorithm puzzles or take-home projects. Candidates complete two live sessions: a fundamentals assessment without AI tools, and an AI-native assessment with whatever tools they would normally use. Across both, assessors look at things like ownership, judgement by risk, and how candidates handle ambiguous tickets rather than well-formed ones. The full scorecard criteria and session format are described at how we vet.
Short answers
Is the coding interview completely dead?
Not completely. Many large companies still use them at scale for filtering. Smaller teams and those hiring for AI-native roles are moving away fastest, replacing them with task-based reviews or live sessions where AI tools are permitted.
What is the best alternative to a coding test?
No single format is best. Most teams combine a realistic take-home task with a structured conversation or walkthrough. The key is measuring judgement and review ability, not memory or puzzle speed.
Can candidates just use AI to pass a take-home coding test?
Yes, which is why the review of the submission matters as much as the output. A follow-up walkthrough, asking the candidate to explain decisions and spot introduced errors, quickly separates genuine understanding from generated code that was accepted uncritically.
