An AI-native technical interview that measures engineering judgment, not AI output.
Candidates receive a realistic product brief and a fixed amount of time to build. They work in an environment that feels like their day-to-day workflow, with an AI coding agent available from the start.
Just like Codex, Cursor, or Claude Code, the agent can propose designs, write implementations, and answers questions. But some suggestions are intentionally flawed, revealing whether candidates verify, challenge, or blindly accept AI output.
We evaluate more than the final project. We track how candidates used the AI, when they questioned its recommendations, what they corrected, and how effectively they balanced speed with sound engineering decisions.
Interviewers receive a concise report covering project progress, AI usage, critical decisions, and demonstrated engineering judgment—making it easier to identify candidates who can work effectively with AI instead of simply relying on it.
AI coding assistants don't understand the nuances of your business. Modern engineers need to direct agents towards the best solution to specific problems, not just generic solutions.
How we prove itWe hand them a live agent which intentionally steers in sub-optimal directions. Do they push back, or ship the first thing it proposes?
AI writes code, and suggests ideas that look right but often aren't. Catching what's confidently wrong before it ships is what separates an engineer from a code generator.
How we prove itOur agent plants subtle, realistic bugs on purpose, and suggest subpar ideas intentionally. Your report shows exactly which they caught and which slipped through. You get a live replay of how it all happened.
Engineers must be accountable for every line that ships, whether they wrote it or the AI did. They must understand the why as much as the what.
How we prove itThey're graded on the final code as their own. Then, at the end, a neutral examiner asks them to defend the decisions they made.
After the build, an AI examiner takes over. It reviews the finished project and generates questions tailored to the candidate's implementation, and your preferences, giving them a chance to explain their decisions and demonstrate their reasoning.
You switched the pagination divisor to Math.ceil. What was wrong with the original — and how would you have caught it if the tests hadn't?
Integer division dropped the last partial page. I'd catch it with a total that isn't a clean multiple of the page size — the count comes back one short.
The retry handler still isn't idempotent. Was leaving that a deliberate call, or did it get past you?
Illustrative exchange — not a real candidate.
Measure both how well they managed the agent, as well as their final output. Supervision is measured against mistakes we planted on purpose. Craft is a read on how they build. See how many of the agent's mistakes they caught, and how well they met the spec they were charged with crafting all in one report.
Defended the pagination fix in their own words. Couldn't justify the missing idempotency check — flagged for the panel.
Illustrative example — not a real candidate.
Sources: US Dept. of Labor (≈30% of first-year earnings) · recruiting/backfill average · $500/day vacancy cost. Adjust the sliders to your own numbers.
The spread between a top and a bottom hire runs ~$106,000 a year in output — every year they stay.