I hire the judgment to catch AI, not the fluency to use it
An agent burned real money being confidently wrong while I was in meetings. That is where my thinking about hiring engineers in the AI era starts, and it does not end where the hype does.
I have never run this interview. Say that first, because the rest sounds like advice and it is closer to a confession.
A few months ago I gave an agent a job and went into meetings. It came back confident that it had fixed the problem. It had not. It kept working. By the time I stopped trusting it, it had spent 60 hours of agent time and roughly $15,000 in model spend, and the bug was exactly where I left it. An IAP misconfiguration. It never doubted itself once. I did not check its output for far too long.
That is the whole lesson. AI does not really make you faster or slower. It makes you confident. In one controlled study in 2025, experienced developers were 19% slower with AI on their own codebases while believing they were 20% faster. The measurement said one thing. The feeling said another. The feeling won.
So "do you use AI" is the wrong interview question. Everyone uses AI. I use it all day, and I have not shipped production code since 2015. The question that matters is smaller. Can you tell when it is wrong?
It is wrong most where being wrong costs the most, and where you cannot delegate the judgment. Accessibility is the clearest case I know. Ask an agent for a form and it returns something that looks right and that no screen reader can complete, confidently, with no idea it failed. The accessible structure a blind person needs is the same structure a machine reads best, and the model still will not produce it, because nobody taught it to care and most developers were never taught either.
That is the signal. Not "did you use the tool." Did you catch what it got wrong, in the place where getting it wrong hurts someone real.
So bring your own agent. I am not checking whether it does your work for you. I want to see what you do when it hands you something plausible and broken. Do you ship it or catch it. Do you delegate the boring 80% and stay on the 20% that runs unattended and cannot fail. Do you reuse a proven pattern or reinvent one because it looks nicer.
None of this retires fundamentals. The opposite. For a senior engineer in 2026, no working fluency with these tools is a real negative signal. Not an automatic no. Strong, broad, current fundamentals can outweigh it. But fluency is a thin layer. Without the deep engineering underneath it, it is a performance. It is not a system that stays up.
I have not solved all of it. A polished setup can be bought or copied. Free tiers are not fluency, and time to learn is not handed out evenly. And I have no clean answer yet for the candidate whose own tools are not accessible to them, which where I work is the least comfortable gap of all.
So, no 10x. No agent as a mirror of the engineer. No claim that the job changed for everyone. Mine changed. Yours is changing on its own terms. What I hire for did not. I do not care what your agent can produce. I care whether you know when to stop it.