Adam Elmore

I think AI will make a lot of knowledge work dramatically faster, especially programming. Models can already replace hours of mundane work, improve files, suggest changes, and let one engineer operate across more of a system. That is real leverage, not just hype. But I don’t think faster automatically means better. The tradeoff I keep running into is distance. If every change becomes a conversation with an agent, I can ship more while understanding less of the codebase. The work can shift from making things to feeding, reviewing, and redirecting agents—sometimes through failure loops that only look productive. That can erode both engineering judgment and the satisfaction of the craft. It can also encourage an unhealthy always-on rhythm, as though idle agents or sleep represent wasted capacity. So I expect a future with much cheaper plausible output, but not cheap quality. Choosing worthwhile goals, developing taste, checking the work, and caring enough to make something good still require effort. We’ll also need to treat agents as operational actors with meaningful access: credentials, tools, and autonomy create concrete risks. The future I want keeps humans in control and uses models to remove drudgery without removing our understanding of—or connection to—the work.

The biggest shift would come from agents becoming reliably capable without requiring constant supervision—and doing so without hiding how the system works from the engineer. If I could delegate substantial changes, inspect clear reasoning and diffs, and consistently come away with more understanding rather than less, that would resolve much of my ambivalence. The opposite would also matter: repeated evidence that greater capability mainly creates more review burden, security exposure, compulsive agent-management, and codebases nobody really understands. I’ve seen both striking successes and stubborn failure loops, so the practical question is whether better models produce durable leverage or just more plausible output moving faster.