What should we learn now?

In corporate software development, large language models have become routine, improving some work and possibly raising the average quality of code. But they also remove the friction through which developers once learned: exploring unfamiliar legacy systems, encountering different approaches, and following technical side paths. As coding agents take on implementation and hide much of their process, developers increasingly supervise several projects and respond to prompts rather than working directly with code. The concern is that this shift turns programmers into managers and gradually erodes the practical skills and understanding they need to do the work themselves.

Because employers have little incentive to sacrifice the speed gains of automation, maintaining those abilities may require deliberate practice outside work. Aviation offers a model: pilots still need hands-on flying experience to stay competent, even when autopilot handles much of a flight. Developers can similarly build projects by hand, learn new languages, study foundational material, and work below the abstraction level they normally use - including writing a coding harness themselves. The argument is not that AI tools are difficult to learn; they are easy to adopt, and users who delay may retain stronger technical skills and fewer assumptions about what models can do. Although some developers can use agents without losing touch with the code, the safer course is not to assume that discipline comes automatically, but to invest in it deliberately.