Hacker News new | past | comments | ask | show | jobs | submit
I share the absolute opposite opinion as your conclusions. I see agents make awful architectural decisions all the time, and I have been seeing the effects of this throughout the boring enterprise Java codebase I work in. We have classes that are thousands of lines, with tests that are also thousands of lines. Every single PR over the past few years in these classes probably seemed reasonable in isolation, but the problem starts to become untenable. I genuinely cannot reason about the data flow in some of these classes.

If a human spent the upfront effort to create and refine the right abstractions, the code would almost always be easier to reason about, test, and extend for agents. I pretty much never see AI doing this on its own—things like proposing to split the 1400 line class into 5 classes, make major changes to the data model, etc. It will most often just make the existing code a little bit worse to answer your prompts and move along.

So I think that the most valuable time spent when working with agents is to ensure that good abstractions are being created, and often this can involve doing this part by hand and asking it to fill out the skeleton. Then you end up with a better codebase that is easier for both humans and agents to work in.

I share the same experience that the quality of code output from agents is highly dependent on their ability to simply follow high quality abstractions already laid out. Putting in a lot of work there pays off dividends later.

By this I mean concrete code design decisions, not just deciding what functional requirements need to be met.

"Every single PR over the past few years in these classes probably seemed reasonable in isolation, but the problem starts to become untenable." - So this untenable situation was created by humans? Or if you mean by AI, well, than I agree that older models created such mess. But try Opus 5.5. In my experience, it's creating pretty good code. Yes, some files can get large. But try this prompt: "Audit the code base for DRY, YAGNI, KISS and create a report that other agents can work through to improve the code base." I'd be surprised if there is something really slowing down humans or agents that you can identify but that wouldn't show up in the report.