Efazati

Hiring

Hiring engineers when AI is in the room

Mohammad Efazati · · 3 min read

I have run thousands of technical interviews over twenty years. For most of that time the hard part was the format: LeetCode puzzles that predicted nothing, take-home tasks you could not trust, whiteboard sessions that rewarded people who had practised whiteboard sessions. Then AI arrived and made every one of those formats worse at once.

The common reaction was to ban it. Lock the browser, forbid the assistant, watch for a second screen. I understand the instinct, but it tests a job that no longer exists. The engineer you hire will use AI on their first morning. If the interview pretends otherwise, you learn how someone performs in a situation they will never be in again.

The answer stopped being the signal

When a model can produce a correct function in four seconds, a correct function tells you very little. What still separates people is everything around it:

  • How they frame the problem before they type anything. Do they read the logs, check what is running, ask what "done" means?
  • What they ask the model, and how specific they are. A vague prompt and a precise one say a lot about how well someone understands the system in front of them.
  • Whether they check the output. The dangerous engineer is not the one who uses AI. It is the one who pastes whatever comes back into production.
  • What they do when it is wrong. Models are confidently wrong often enough that recovering from a bad suggestion is now a core skill.

None of that shows up in the final answer. All of it shows up in the process.

Put them on a real machine

The other half of the problem is the environment. Most interview tools give a candidate an editor and a "run" button. Real work happens on a Linux box with a broken service, a full disk, a misconfigured proxy and a log file that is lying to you. Infrastructure and backend roles especially are about the system, not the snippet.

So the interview should look like the job: a real machine, the tools the role actually uses, AI available, and a problem that has more than one reasonable way in. Then you watch how they get there.

Let candidates use AI the way they will at work, and make the process the thing you evaluate.

What I look for now

When I review a session today, these are the moments I go back to:

  1. The first two minutes. Does the candidate orient themselves (systemctl status, journalctl, df -h, reading the task twice) or start changing things?
  2. The first wrong turn. Everyone takes one. How long before they notice, and what makes them notice?
  3. The handoff to AI. Is the model a tool they direct or a crutch they follow?
  4. The verification. Did they prove the fix, or did the error just go away?

A candidate who takes longer but verifies every step is usually a better hire than one who finishes fast and cannot say why it works.

Why I built EasyEnv

I wanted this kind of interview to be the easy option rather than a special effort. That is what EasyEnv does: interviews on real Linux machines in the browser, AI allowed, and a full replay of every terminal, editor and browser action so the team can review the process, not just the result.

If you are rethinking how your team hires engineers, I would like to hear what you are trying.