Raising the floor, lowering the ceiling

A developer reflects that AI coding agents like Claude have broken the link between output and improvement, making it easy to produce large amounts of code without the deep learning that comes from manual problem-solving. The author warns that future developers who start as 'managers of agents' may cap their skill ceiling at the quality of the best available agent, trading the path to better software for mere efficiency.

Raising the floor, lowering the ceiling I find that it is now very easy to write a lot of software. In the past, if I was writing this much software I would take that as a strong signal that I was become a better developer. Now, I write more software than ever but the link between output and improvement is broken. There is a feeling I get when working on a hard problem and writing the code by hand. Taking the thing I need to do and distilling it into code was the work. Seeing the computer do the thing I wanted it to do was the reward. I might try a dozen failed ways to solve a problem. Learning a bit more every time I got stuck or the thing I tried didn't work. I don't get that feeling when working with agents. I output a lot more and there's still a buzz building something that works but most interaction with Claude is abstract from the code in a way that makes it feel different. If the output is wrong, the next step is to say "this output is wrong" and let it try again. Not a bad feeling, just a different one. In a lot of ways its the same feeling I felt first moving from being an IC to a manager. I might read the code sometimes, helping out in very rare circumstances but most of my job became guidance and being a sounding board. Both having a 'team' and having 'agents' opens the door for producing a lot more raw output but in both cases you're ceiling is defined by someone else. The ICs in your team or the agent you happen to have access to. I am a little worried that if future developers start as 'managers of agents' on day one we're accepting that our code is only going to be as good as the best agent we can buy. Time spent writing software isn't a path to writing better software, only a path to writing more of it, perhaps more efficiently.