Why Can Programmers Be Worse at Directing AI Agents?

David Heinemeier Hansson says that for a while his own programming experience was a deficit. He told agents to implement the path he would have taken, and they were very good at obeying. In the first agentic stretch, about three months, that felt productive. He used everything he knew to prescribe the work. He was late to the next stretch, where describing the outcome and the problem produced a better solution than a programmer's prescribed path.

He will say it outright: there are problems for which programmers are worse than non-programmers. A lot of programmers are not very good product managers, and software is product management. What should it do, for whom, how should it look, what ships in version one? Those skills are not evenly spread across people who can write code. Once an agent does the implementation, those are the skills that decide the result. Knowing loops and good engineering still changes how someone directs an agent. Dictating the old path can make the answer worse.