The Interview Loop Is Out Of Date
Most senior engineering interview loops were designed when the cost of writing a working CRUD endpoint was nontrivial. Today, an AI agent writes it in two minutes. The loop that tested whether the candidate could write the endpoint is no longer testing what it thinks it is testing.
This is not a hypothetical. Teams that have rolled out AI coding agents are noticing that engineers who pass the legacy interview do not always thrive in the new environment, and engineers who struggled in the legacy interview sometimes excel. Something has shifted.
This post is about what the new signals are.
What Still Matters
Three signals that became more important, not less:
Code review judgment. When the AI writes the diff, the engineer reviews the diff. The skill of reading code critically, spotting the subtle bug, and understanding the system-level implications of a change has moved from "good to have" to "the core of the job." Interview for it. Show candidates AI-generated diffs and ask them to find what is wrong.
Problem decomposition. The AI is best at well-scoped, well-decomposed problems. The engineer's job is to do the decomposition. A candidate who can take "users are complaining about checkout" and turn it into four specific, testable changes is more valuable than they were five years ago.
Systems thinking. Knowing how a change in one service ripples through three others. The AI can write the change correctly in the local context and still produce a bad system-level outcome. Engineers who hold the whole system in their head catch what the AI does not.
What Matters Less
Three signals that have lost weight:
The whiteboard algorithms grind. Reversing a linked list under time pressure used to be a proxy for "can write code." It is no longer a useful proxy. The AI writes the linked list reversal. The engineer chooses which data structure to use and why.
Speed of typing working code. AI agents type faster. The advantage of being a very fast coder has compressed. The advantage of being a thoughtful coder has expanded.
Memorization of framework APIs. "Do you know what React.useReducer's signature is?" was a knowledge proxy. The AI knows. The engineer needs to know when to reach for it, which is a different question.
New Signals Worth Adding
Three interview moves we have seen work:
The AI-paired review. Give the candidate an AI-generated PR (200 lines, multi-file, in a familiar codebase), and ask them to do the review. What do they catch? What do they miss? How do they communicate with the AI (or the engineer behind it) about the issues?
The disambiguation interview. Give the candidate a deliberately vague problem statement. Watch them ask clarifying questions. The skill of disambiguating before coding is the skill of using an AI agent well. (Also the skill of being a senior engineer in any era, but more central now.)
The escalation call. Give a scenario: the AI has produced a PR that looks reasonable but smells wrong. Walk me through what you do. Strong candidates trace back through the AI's reasoning, identify the missing context, and decide whether to retry with more context or take it manually. Weak candidates either accept the PR uncritically or reject it without diagnosis.
What To Stop Asking
We retired three categories of question:
- Memorized algorithms: no longer predictive.
- Trivia about framework internals: easily looked up.
- "Implement this small thing in 45 minutes": the AI does it in two minutes. Either the test is too easy, or it is testing the candidate against the AI, which they will lose.
The Junior Engineer Question
A real concern: if the AI writes the entry-level work, how do you train new engineers? Three things we have found:
Junior engineers as reviewers, earlier. They review AI-generated PRs. The exposure to read-then-evaluate is faster than the legacy path where they wrote everything themselves. They become senior-quality reviewers a year earlier.
Junior engineers as scope shapers. They sit with PMs, learn to disambiguate, and produce the scoped tickets the AI executes. The product-engineering skill they develop is the senior skill, just expressed in a new format.
Bigger projects, sooner. With the AI handling the implementation grind, junior engineers can be responsible for the architecture of a small service in their first year. The career growth slope is steeper, not shallower.
What Hiring Managers Should Push On
Three things to demand when interviewing for an AI-heavy team:
Have they reviewed AI-generated code? If they have not, they need to ramp up before they will be effective. Ask for specific examples, what they caught, what they missed.
Do they have opinions about agent behavior? A candidate who has formed opinions ("the AI should refuse vague tickets," "I do not trust the AI on cross-service changes") has done the thinking that lets them work with the AI productively.
Can they articulate where the AI is useful and where it is not? Candidates who think AI is universally great are naive. Candidates who think it is universally bad are stuck. Candidates with a calibrated view are useful.
What We Got Wrong
Two things we initially over-emphasized:
LLM prompt engineering as a skill. We tested for it early. It turns out to be a low-stakes skill that anyone picks up in a week of working with the tool. It is not a hiring filter.
AI tool stack familiarity. We thought knowing GitHub Copilot vs. Cursor vs. Continue mattered. It does not. Engineers learn the new tool in days. The underlying judgment skills are what we should have been screening for from the start.
The Quiet Implication
Hiring engineers in the age of AI coding agents looks more like hiring senior engineers always did (for judgment, taste, and systems thinking), and less like hiring junior engineers used to, for productivity and pattern matching. The bar has moved up, in the sense that everyone is expected to do work that used to be reserved for seniors. It has also moved differently, toward skills that some excellent engineers were undervalued for in the old loop.
The teams that adjust their loop catch better engineers. The teams that do not, hire for the wrong signal and wonder why the new hires are not productive.
For the broader rollout context, see engineering manager playbook. For why this changes what reviewers do, see the reviewer's toolkit.
Frequently asked questions
What should you look for when hiring engineers who work with AI coding agents?
Interview for the skills that got more valuable, not less: code review judgment (reading a diff critically and spotting the subtle bug), problem decomposition (turning 'users are complaining about checkout' into four specific testable changes), and systems thinking (knowing how a local change ripples through other services). These are the skills that determine whether someone thrives on an AI-heavy team, because the agent is best at well-scoped work and the engineer supplies the scoping and the review.
Are whiteboard coding interviews still useful in the age of AI?
Largely no. Reversing a linked list under time pressure was a proxy for 'can write code,' but the AI writes the reversal now, so it no longer predicts on-the-job performance. Speed-of-typing and framework-internals trivia have similarly lost weight. A 45-minute implement-this-small-thing exercise is either too easy or tests the candidate against the AI, which they lose, retire it in favor of review and disambiguation exercises.
How do you train junior engineers when AI writes the entry-level work?
Change what juniors do rather than assuming there is nothing left. Put them on AI-generated PRs as reviewers early, which builds senior-quality review judgment about a year faster than the write-everything path. Have them shape scope alongside PMs, learning the disambiguation skill that is really the senior skill in a new format. With the agent handling the implementation grind, juniors can own the architecture of a small service in year one, a steeper growth slope, not a shallower one.
What interview questions test a candidate's AI code review skills?
Give the candidate a 200-line, multi-file AI-generated PR in a familiar codebase and ask them to review it, watch what they catch, what they miss, and how they communicate the issues. Follow with an escalation scenario: an AI PR that looks reasonable but smells wrong, and ask them to walk through what they do. Strong candidates trace back through the AI's reasoning, identify the missing context, and decide whether to retry with more context or take it manually. See the reviewer's toolkit.
Is prompt engineering a skill worth hiring for?
No. Prompt engineering turns out to be a low-stakes skill that anyone picks up within a week of working with the tool, so it is not a useful hiring filter. The same is true of AI tool-stack familiarity, whether a candidate knows one autocomplete or agent tool versus another does not matter, because engineers learn the new tool in days. Screen for durable judgment: code review, decomposition, and systems thinking. For rollout context, see the engineering manager's playbook.
EnsureFix Engineering Team
The EnsureFix engineering team designs and operates the multi-agent pipeline that turns tickets into production-ready pull requests. They write about architecture, model routing, safety validation, and what actually ships in enterprise codebases.
