Hiring in the AI Era

When anyone can produce plausible code with an agent, the coding screen stops predicting much. Hiring shifts toward what agents don't supply: judgment, systems thinking, debugging under uncertainty, communication, and the taste to tell good output from merely plausible output.

The Pattern

When anyone can produce plausible code with an agent, the traditional closed-book coding screen stops predicting much -- it tests the thing that is now automated. Hiring shifts toward what agents do not supply: judgment, systems thinking, debugging under uncertainty, clear communication, and the taste to tell genuinely good output from output that merely looks right.

Large engineering orgs are already redesigning the screen itself rather than policing AI out of it. In October 2025 Meta began rolling out an "AI-enabled coding" interview; an internal message described it as "a new type of coding interview in which candidates have access to an AI assistant. This is more representative of the developer environment that our future employees will work in, and also makes LLM-based cheating less effective" (Hello Interview, 2025). Candidates work inside a multi-file codebase they did not write, choosing among models such as Claude Sonnet, GPT-5, and Gemini, and are graded on the same four competencies as the classic format -- problem solving, code quality, verification, and communication -- now expressed through how they direct the tool. As one internal Meta source put it: "Should use AI, but need to show you understand the code. Explain the output. Test before using. Don't prompt your way out of it."

Why It Matters

Screening for syntax recall now selects for exactly the skill agents have commoditized, while missing the skills that determine whether someone can direct and check an agent's work. The reframing is subtle: making AI available does not make the interview easier. Reported experiences range from a candidate finishing all three phases in 40 minutes to an E7 who watched a strong model "repeatedly hallucinate on a maze problem" and abandoned it to write the code himself (Hello Interview, 2025). Weak fundamentals surface faster, not slower -- "AI can actually make weak engineering fundamentals more visible," because gaps appear the moment a constraint changes or an edge case is probed (ul Haq, 2026). Verification and the ability to spot AI failure modes -- off-by-one errors, wrong complexity, hallucinated APIs -- become the signal, which connects directly to comprehension debt and the move from coder to orchestrator.

The honest tension: these formats are early, narrow, and unproven at scale. Meta's is still an invite-only pilot drawing from a small problem pool, runs alongside a traditional no-AI round, and several candidates report the in-interview model is quietly "nerfed" via system prompt versus the practice environment -- so the signal is partly about coping with a deliberately weakened tool, not real-world conditions. There is also a deeper measurement problem the interview does not touch: as titles rise, code volume stops being the job at all. One self-described "ghost engineer" notes that when a large project is off track "it's rarely for a lack of coding velocity -- it's a lack of consensus, direction, customer buy-in, system architecture," and "none of the artifacts for those issues get committed to GitHub" (Gengelbach, 2024). The durable hiring signals -- can this person think in systems, find the bug nobody else can, and judge what "correct" means -- are also what performance will later be measured on, and no single interview format has yet shown it can reliably capture them.

Last reviewed: 2026-06-25