The way we build and ship products has changed quite a bit over the last year. A couple of years ago, we were pasting error messages into this new tool called ChatGPT to fix the errors we ran into. Time passed and we started asking it to build features, paste the snippets it generated into our codebases and hope everything would fall into place.
Well, we've come a long way since then. New tools launched, new workflows were unlocked and a whole new way of coding emerged: vibe coding agentic development.
I don't remember the last time I wrote more than 10 lines of code in one go, and why would I need to? My Codex agent takes care of my tasks, writes test cases, uses my computer to take a few screenshots, validates the results and fixes any issues it runs into.
More than ever, focus is at an all-time low as my agent just quietly does the work while I browse stuff online, only for me to take a quick glance at the hundreds of lines it generated and assume everything's fine.
The thing that bothered me wasn't that the agent got something wrong. It was that I wasn't particularly interested in finding out whether it was wrong.
A couple of years ago, if I wrote a feature myself, I would obsess over every little detail. I'd read the code multiple times, think about edge cases, refactor things that nobody would ever notice and generally spend far too much time polishing it. Today, I find myself staring at a diff generated by an agent and wanting to hit approve as quickly as possible.
Part of it is that the models are genuinely getting better. Part of it is that the tooling around them is getting better. The agent runs tests, takes screenshots, validates the output and fixes obvious issues before I even look at it. By the time the work reaches me, it's often good enough.
But "good enough" creates a strange problem. The less often the agent is wrong, the less motivated I am to verify that it's right.
Maybe that's just the natural progression of the job. Less time producing, more time reviewing. Less individual contribution, more delegation. In some ways it feels less like programming and more like management.
I don't know if this is temporary or if this is simply what software engineering looks like now. What I do know is that writing code was never the difficult part for me. Paying attention was.
And as the cost of generating code approaches zero, attention increasingly feels like the thing in shortest supply.