Opinion. Written by an engineer who still checks which way the stack grows, just in case.
I started at the bottom of the stack: bus clocks, RAS and CAS signals. Software was what happened after the memory controller said yes.
Then vibe coding arrived. I watched an agent type a whole service while I sipped coffee, and a 2 a.m. thought followed: would I choose this job again?
I would. It took a cache line and a few thousand job ads to be sure.
The Doubt Is Real, and the Data Shares It
"Vibe coding" is Andrej Karpathy's name, from February 2025, for describing what you want and accepting what the AI writes. Fine for a weekend project. Less fine in a high-frequency trading engine, where a bug gets filled at market price before you can read the log line.
Ask Knight Capital. On 1 August 2012, one of its eight servers still ran old code. In the first 45 minutes of trading it sent more than 4 million orders while trying to fill 212, and lost more than $460 million (SEC). No AI involved, just a deploy that was almost right.
The 2025 Stack Overflow Developer Survey asked developers how it goes:
- 84% use or plan to use AI tools. Up from 76% a year earlier.
- 46% distrust their accuracy. Only 33% trust it.
- 66% say the answers are "almost right, but not quite". The most common complaint.
- 45% say debugging AI-generated code takes longer. The time saved on typing went into debugging.
A METR study is even more humbling. Sixteen experienced open-source developers worked on 246 real tasks. With AI, they took 19% longer. They had expected to be 24% faster.
That's a small study of early-2025 tools, and the tools got better since. Still, it matches what I see: typing got faster, and typing was never the slow part.
"Almost Right" Is the Puzzle Factory
This is what cured my 2 a.m. doubt.
Wrong code is easy: it crashes, you fix it. Almost right code is the expensive kind. It passes the tests, ships on a Friday, and fails on the one input nobody imagined.
AI produces almost right code at scale. Someone has to find the "almost", and that takes knowing how the machine really works.
So the work moved. Writing code is cheap now. Checking it is the job.
What the Job Ads Say
We run a job site, so we checked. We analysed 9,425 Software Engineer postings from 3,128 companies, published from March to September 2026.
- It's the most-posted AI-related role. First of the 44 roles we track, ahead of Project Manager with 6,539 postings.
- AI is mostly a tool here. AI Assisted Coding appears in 65% of postings. Building AI is the job itself in only 8%.
- The missing skills are about checking. The AI skills we most often flag as missing: AI Security Awareness (23% of postings), Prompt Engineering (22%), AI Code Review (17%) and AI Testing Automation (14%).
Three of those four are security, review and testing. Employers can find people who get an agent to write code. They're short of people who can tell whether that code is safe, whether it's right, and how to prove it.
The survey and the job ads agree: checking is the work now.
A Few Puzzles for Your Curious Mind
Checking is more fun than it sounds. Try these before asking a chatbot (I won't tell).
The old ones
- Does the stack grow up or down? On x86 and ARM, down; on PA-RISC, up. The ABI decides. Now prove it in C, and find out why inlining and undefined behaviour can make your proof lie.
- Two threads, two counters, no shared data. Why so slow? Both counters sit in one 64-byte cache line, and the cores fight over it. The fix is a few bytes of padding.
- The bug that vanishes when you add a printf. Usually undefined behaviour or a race. The printf changes the timing, the memory layout or what the optimiser does, so the bug politely leaves the room.
- Redis is "just a cache". Until a restart empties it and the database meets real traffic for the first time.
Look at what these have in common. The machine did exactly what it was told, and someone was still surprised. The answer sat one layer below the code: in the ABI, the cache line, the restart.
That moment when it clicks is why I do this job.
Agents are a new layer with the same kind of surprises. Squint and the puzzles look familiar:
The new ones
- The context window is a cache. Something gets evicted. Which instruction did the agent forget, and when?
- Prompt injection is SQL injection in English. A web page tells your agent to ignore its instructions. How do you sanitise a sentence? You can't. You limit what the agent is allowed to do.
- A function that answers differently each time. Unit tests assume determinism. It's the printf bug again: same input, different result, and the cause is somewhere you aren't looking.
- The agent's diff is 2,000 lines and the tests are green. What do you read first?
If you enjoyed the first list, you already have the mindset for the second. And agents make more of these every week, because there's more code to be almost right.
So, Would I Choose It Again?
Yes, and the agents are part of the reason.
The boring part got automated: boilerplate, glue code, the fourth CRUD endpoint this week. What's left is the part that made me pick this job: why does this machine do what it does?
I still get to dig into what's under the abstraction. Now I also get to work above it: I can run several agents on one system at once, and spend my time on the design instead of the typing.
Both ends of the stack got more interesting.
(If you never stared at a timing diagram for fun, ask an agent which way the stack grows. Then check.)
Start This Week
I read every agent diff as a puzzle now. Somewhere in it is the "almost", and finding it is my job.
Try it this week. Treat the next piece of AI code as a puzzle, not as finished work:
- Pick one puzzle from the list. Solve it without AI first, then compare with what the AI says.
- Read one AI diff line by line. Find the "almost". There usually is one.
- Go one layer down. If you write Python, read how a dict works. If you write C++, read what your CPU does with a cache miss.
- See what employers ask for. Our Software Engineer guide lists the skills and gaps, and the AI jobs in the Netherlands show who is hiring.
No bomb went off. I got better puzzles.
Figures: SlashHash analysis of 9,425 Software Engineer job postings from 3,128 companies, March to September 2026, from our Software Engineer AI career guide. Stack Overflow, 2025 Developer Survey, July 2025. SEC, Knight Capital charges, October 2013. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, July 2025.