AI writes most of my code now. The architecture is still mine.

After 30 years of building software, my job has quietly changed shape this year.

The routine work - the structure, the formatting, the scaffolding that used to eat hours while the actual problem waited - I hand that to an agent now. I describe the task, it works across the codebase, runs the tests, and comes back with a result. My part looks more like reviewing a very fast junior developer with infinite confidence.

That frees my attention for what was always the real job: the logic, the purpose, the decisions.

But the tool has taught me a few things the hard way. AI amplifies the direction I give it - I don’t rely on it to tell me when that direction is wrong. A precise task produces impressive work at remarkable speed. An imprecise one produces impressive-looking work that is confidently, fluently wrong. Nowhere is that gap more expensive than in security, where code that looks right and code that is right are two different things. Its confidence isn’t evidence. I still have to verify the result.

This is where the 30 years turn out to matter more than I expected. Experience doesn’t make AI unnecessary - it tells me what to question, and what still needs testing. I’ve spent decades learning what software costs after day 1: maintenance, integrations, security, the day the person who built it leaves. AI has made day 1 dramatically cheaper. It hasn’t removed the need to think about day 500.

Once in a while, something I was sure would work one way turns out to work the opposite way. That used to be a bug. Now it’s often the most interesting moment of the week - the machine showing me an assumption I didn’t know I had.

The tool is extraordinary. The judgment is still the job.