Hiring junior engineers when AI writes the code

When Copilot and generative AI tools became ubiquitous, engineering managers feared that evaluating junior engineers would become impossible. If an applicant can generate a complete full-stack app with a single prompt, what does a take-home assignment actually measure? Over the past two years, placing thousands of developers in engineering organizations worldwide, we noticed the opposite: AI did not obscure talent, it polarized it. Junior candidates now fall into two distinct groups: 1. **The Assembers**: They copy prompts, glue libraries together, and panic when an unhandled runtime exception appears. They cannot explain the state lifecycle of their own application. 2. **The Systems Thinkers**: They use AI as an accelerator for syntax, but maintain complete mental models of their data flow, architecture boundaries, and failure modes. In our placement interviews, we stopped asking candidates to write boilerplate from scratch. Instead, we give them a running application with a subtle, race-condition bug and ask them to debug it live with access to AI tools. The candidates who succeed are those with high **debugging stamina** — the ability to form a hypothesis, instrument logs, inspect network payloads, and verify assumptions systematically.