some basic thoughts on the discourse between mario/cramer/dax and i think a few other ppl on twitter: i definitely agree, was thinking this near the end of 2025 but i just thought i wasn't doing it right lol
i think the thing that's hard to explain to people (including my own boss) is that, no matter how well you plan or review or even understand the agent's code, the code is still "slop." slop in that there inevitably decisions in there (even decisions that are as trivial as where to put a function boundary) made by the clanker. and since there were decisions made by the clanker, some of those decisions are inevitably bad. trivial decisions like the structure of an interface may work now but will make adding features and harder. the thing that is hard to explain is that tech debt is 1) almost impossible to undo 2) makes developing software significantly harder 3) is an inevitable result of agent-driven development.
as for how to fix it, i have no idea. the idea i've been pursuing lately is lowering the level of decisions human make. for example, i write the interfaces, i write the outline of a class that implements that interface, i write methods on the interface. then i ask pi to implement the methods. the methods have a very clearly defined purpose and really only one "right" way to have it implemented (and the clankers are pretty good at finding the "right" way) so a brief review of the clanker's code is enough. of course, this is significantly slower than having an agent write everything, which is also what makes it hard to convince people to switch back.
Recent Announcements
Continue the conversation
Join the Discord to ask follow-up questions and connect with the community