How we vet AI-native engineers

Most technical interviews still optimize for the 2018 engineer: can you whiteboard a graph traversal, can you debug a closure, can you talk through system design. None of that screens for the skill that actually predicts on-the-job AI-native productivity.

Here's the loop we run instead. Four stages, total candidate time ~3.5 hours, and we get a higher hit-rate than the old eight-round process.

The take-home is the single biggest signal. We don't grade the code — we grade the process: did they decompose the problem before coding, did they verify the agent's output, did they catch the subtle bug the LLM introduced. Engineers who treat the agent as a colleague (review, push back, test) sail through. Engineers who copy-paste whatever appears struggle visibly in the live pairing.

The architecture chat is where we test the orchestration muscle from "One engineer, three agents". Can the candidate take a fuzzy "build a notification system" prompt and split it into three well-scoped sub-tasks that could be parallelized? That's the headline skill. Everything else is downstream.

Worth noting: this loop is harder to game than leetcode. There's no LeetCode for "decompose ambiguity well in front of a senior engineer." The candidates who pass are the ones whose day jobs already look like this.