Hiring an AI developer in 2026 means filtering out the fakes first.

Hiring a dedicated AI developer in 2026 is hard. I’ve been on both sides of it, so let me start with the candidate side, because that’s where it gets weird first.

You can have a really good career. Ten-plus years, a bunch of companies, projects you actually led, not just worked on. A GitHub that isn’t empty, a blog, some followers on LinkedIn and Twitter. And you still end up in this completely disconnected pipeline where you apply, nothing happens, you apply again, nothing happens. Because the queue on the other side is flooded with AI-generated resumes. Most of them aren’t relevant to the job. They look relevant. That’s enough to push you out.

Now I’m on the other side, hiring. I open the applicant list and it’s two hundred people who are all senior, all “LLM”, all “agents”, all “production”. I know most of it isn’t true. I just don’t know which parts. And I’ve got a few weeks and a handful of interview slots before the project starts slipping. So this post is me trying to write down what I actually do about it.

The resume that learned Go overnight

A typical one: the candidate’s real background is PHP. The role is Go. They’ve got maybe a few months of Go, if that. But the resume went through a model and now they’re “super proficient in Go.” And it’s not just Go. Kubernetes, RAG, agents, whatever was in the job post is in the resume now. Do that across hundreds of applicants and the whole market inflates. Every resume says senior and every resume matches the keywords.

Half the time it isn’t even deliberate lying. People paste the job description into ChatGPT, type “tailor my resume to this”, and the model does what it’s told. Three weeks of poking at a Go service becomes “led the migration of core services to Go.” One notebook that called an embeddings API becomes “designed RAG pipelines for production search.” Every line is loosely connected to something real. The person the resume describes doesn’t exist.

So what happens to the actual experienced developer? They barely get to the recruiter call. Their resume says what they did, in their own words, and that matches the keywords worse than a resume that was written to match the keywords. I know a lot of good people who basically gave up on applying and just ask around - old colleagues, friends, the network. The front door doesn’t work for them anymore.

Play with the slider. The seniors don’t get any worse. Same skills, same resumes. The pool around them changes, and a keyword screen can only count keywords, so at some point the honest resume just drops off page one and nobody ever opens it. People like to say “good engineers always find work.” Sure, eventually, through a friend. Just not through your job post.

Everyone uses AI, nobody admits it

And it’s disgusting, because it’s broken on both ends at once. AI filters the candidates. AI writes the resumes. Everybody pretends they’re not using it and talks about wanting a “genuine connection” with applicants. Meanwhile it’s one model writing a document for another model to read, with two people on either side acting like it’s a conversation.

Then there’s who reads your resume when a human finally does. Usually someone with zero idea what your GitHub project is. They might not know the difference between chat and rank. But they want fifteen years of Python, and it doesn’t matter to them that your Python was PyTorch serving and the other guy’s was Pygame. I’ve seen those two get confused. More than once.

If you’re hiring, fix this first, it’s the cheapest part. Whoever does the first pass has to know enough to ask “which kind?” Which Python. Which production. Which agent - a chatbot with one tool, or something that plans, calls tools, and has to decide on its own when to stop. If your screener can’t ask that, you’ve basically got a keyword counter with a salary, and the tailored resumes will beat it every single time.

The guy who never owned anything

Hiring side again. You can end up interviewing someone who’s using AI to answer your questions live, on the call. Annoying, but that’s not the bigger problem. The bigger problem is what happens after you hire them. Someone who looks like they’re working but never actually owns anything. Low intent, no real interest in your project. They wanted a job. They’ll sit there for six months until you let them go, and they’re fine with that.

I’m not ranting, this is just the market right now. Some people got so good at gaming the resume process that they hold two or three jobs at once and juggle between them. Which makes hiring even harder, unless you’ve got trust, a reference, something that tells you this person has actually been near production.

The live-AI thing is easier to catch than you’d think, as long as you stop asking textbook questions. A model will nail “what’s the difference between fine-tuning and RAG.” It struggles with “on the project from your resume, why that chunk size, and what did you try first?” Go deep on their own stuff. Then go one level deeper than feels polite. Change a constraint in the middle - “ok, now you have a quarter of the context window, what breaks first?” People who built it answer fast and kind of messy, with details nobody would bother inventing. People reading off a second screen go quiet for a few seconds and then give you something clean, structured and generic that never mentions their own system.

The multiple-jobs people you usually don’t catch in the interview. You catch them in the first few weeks. They’re slow to reply during the hours you agreed on. Camera’s always off. Work shows up in bursts right before a check-in. PRs look fine, but when you ask about a specific change they can’t really explain it. Any one of those can be nothing, and I wouldn’t accuse anyone over one of them. That’s why I set up the first weeks so I can actually see this stuff, instead of figuring it out in month five.

“AI production experience” usually isn’t

This one I see constantly. People say they have AI production experience and they’ve never run an LLM in production. Never even put a basic PyTorch model behind an endpoint. So they have no feel for how hard it is to go from a prototype to something that runs in a real system with real users and real, messy data.

LLMs are different from classic ML, from computer vision, from traditional inference. But it’s still hard, mostly because there’s a ton of data engineering in it. Retrieval pulls the wrong chunk. The agent loops on the same tool call. Nobody cleaned the inputs. The bill triples overnight and nobody knows why. I wrote about some of that in RAG beyond the tutorial and the tool-call cycles post. You either learn this stuff by breaking things yourself, or you hire someone who’s been breaking things for years on real projects.

Otherwise you get someone who finished the happy-path tutorial. That tutorial covers maybe 10% of what you’ll see in production. The other 90% is where your project actually lives or dies, and a resume can’t show you any of it.

That table is pretty much my interview for the agent part. I don’t want it recited back to me. I want to hear that they ran into each of those walls themselves. I ask how their agent knew it was done. If the answer is “the model says it’s done”, they haven’t run it long enough to watch it loop forever. I ask how they knew a change made things better and not worse. No eval set means it was vibes. I ask what they looked at when a user said “it gave me a weird answer yesterday”. No traces, no way they found it. None of this is a trick. Anyone who’s owned an agent in production for a few months has an answer for all three, and usually a story with it.

What still works

So yes, you can still hire a good AI developer. You just have to do the filtering yourself now, because the pipeline won’t. What’s left are the signals AI can’t fake cheaply.

References are useless if you ask “were they good?” You’ll get “yeah, great to work with” every time. Ask what they owned end to end. Ask what happened the last time it broke. Ask “would you put them on your most important project tomorrow?” and listen to how long they pause before answering. And find a couple of people yourself - someone from their commit history, a co-maintainer on one of their repos. Not just the two names the candidate handed you.

Shipped stuff is the other half. GitHub, a blog, a talk, some open-source tool people actually use. Stars don’t matter much. What matters is that you can open it. You can read the commits and see if the hard parts were written by the same person who wrote the README. You can see how they talk to strangers in issues. That’s months of real behavior and you get it for free. A resume can’t compete with that.

The failure story is the one I lean on most. Someone who ran a model in production will tell you about the night the embeddings drifted, or the retry loop that ate the budget, and throw in details you didn’t ask for. Someone who read about it gives you the tutorial answer. And for the interview itself, stop making people write code, a model writes it in seconds. Have them review it instead.

So

It’s getting harder to find these people in 2026. The pipeline is mostly noise, both sides use AI and pretend they don’t, and the keyword screen is actively pushing the good ones out. But they’re out there. Mostly they’re not the ones winning the keyword lottery. They’re one reference away from you, and they’ll happily tell you about the worst production incident they ever caused. Hire that person.

If you’re looking for one right now, here’s what I do.