While I was reasonably confident in the quality of that previous coding session for tagging, it turns out that it left a lot to be desired.
First off, looking harder at the queries it came up with for hitting the database, one required a full table scan every time, another needed an index to avoid that. The "you're prompting wrong" crowd would argue that you should have told it that, but this is basic wisdom that a developer would have early in their career. It's not even something I'd expect to be specified in requirements, because it's an implementation detail that devs just know.
Then there's the weird stuff it does, like invent extension methods for things that already exist. It seems that the bigger a system gets, the more the robot tries to build things that already exist. I find that I often need to point it at things as a starting point so it doesn't do that, which requires that the human still knows and understands the system. That's really at odds with the tech bros who think they can ship a big system and expect it to scale and be maintained.
So the refactoring and testing, which found a lot of problems, took another two hours, bringing the entire time investment to about five and a half hours. That's pretty close to the estimate I made about me doing it myself. Granted, that's total working time, not accounting for ADHD side quests. Using AI agents is absolutely fantastic, but there's a lot of nuance in the actual benefit involved.
No comments yet.