On doing "deeper work" with AI
For long stretches of 2025 and 2026, most of my work involved some combination of prototyping, very well-known implementation patters, and working in codebases I already knew well. I now have more experience doing (i) production-stage work (ii) in AI-developed parts of codebases (iii) on tricky, details-matter, think-about-them-in-the-shower types of problems.
Here are some impressionistic reflections about this:
- Yes, it feels strange.
- Yes, there is plenty of tech debt to pay down, much of which would not exist had I hand-coded it.
- But there's also tech debt that doesn't exist that would exist if I had hand-coded it. So, for example, sometimes AI will use a slightly less familiar, but still obviously correct, data structure or algorithm, and this will make a refactor trivial. I very much default to simple algorithms, but my sense of what is simple is bounded in part by my knowledge and experience. The AI knows more simple implementations.
- Paying down AI tech debt for hours on end is a grind, and a new kind of grind.
- But the ratio of remediation effort to previous progress is excellent, largely because I have AI to help me in remediating things. This is an area where juggling many AI instances is efficient. ("What the heck is this?" "OK, what the heck is this?" "Can I split this up into X and Y? How many callers would be affected?" And so on.) It is no indictment of agentic coding if one occasionally needs to spend 10 or 20 hours wrapping one's mind around code that was developed months and months faster than it otherwise would have been. (But it might not always feel this way at hour 14 of the process.)
- I disagree strongly with the (common) idea that one should not use AI for the fiddly, creative, tough parts of the job. Yes, my inputs here are indispensable (phew!), but I'd hate to lose the AI's help. It knows the literature; it's good at critiquing proposed designs; it has seen many things. This nice Antirez piece that has been making the rounds is on point.
Many parts of the job are hard, and some are hard in new and uncomfortable ways. But I'm happy to double down on everything I've said about "cognitive debt" here, here, and here.