Merge is shifting engineering hiring from writing code to

LeetCode is becoming irrelevant because most engineers aren't manually writing every line of code anymore; they're managing AI-generated PRs. When PR volumes triple but the actual "typing" decreases, the real skill isn't syntax—it's judgment. The hardest part of the modern AI workflow isn't reviewing your own AI's output, but auditing code generated by another person's AI when you have zero context.

This is the gap Merge is trying to fill. Instead of asking a candidate to solve a puzzle from scratch, it tests their ability to perform a high-stakes code review, which is where senior engineering talent actually spends their time now.

How the assessment works

The process is designed to simulate a real-world asynchronous collaboration loop rather than a static test:

1. The candidate gets a small codebase for context and a specific Pull Request (PR) to review.
2. They leave comments on the PR, pointing out bugs, architectural flaws, or vulnerabilities.
3. An AI agent acts as the submitting engineer, responding to those comments either by arguing its case or pushing a code fix.
4. This loop continues until the candidate hits a limit of five revisions or the timer expires.

The evaluation metrics

Since this isn't a "pass/fail" unit test, the scoring focuses on engineering judgment and operational efficiency:

• Coverage: Did the candidate actually find the hidden bugs or security holes, or did they just nitpick the formatting?
• Communication: Is the feedback constructive and professional, or is it vague and inefficient?
• Efficiency: This is the most interesting part. Merge tracks how many revisions and tokens were consumed during the process.

Tracking token usage and revision counts as a hiring metric is a smart move. In a real-world AI workflow, an engineer who can guide an LLM to the correct solution in one prompt is infinitely more valuable than one who spends ten prompts looping through the same error. It's a direct proxy for how much they'll cost the company in LLM API spend and wasted engineering hours.

For anyone building an AI workflow or trying to implement a more practical tutorial for onboarding new devs, this approach to "review-first" evaluation makes a lot of sense. It moves the needle from "can this person code" to "can this person lead a technical project using AI agents."