It depends. At least, I find that it has a habit of falling on its metaphorical face if the task is anything more complex than the simplest things, so the idea that people can use it to make viable programs is baffling to me.

"Put these values into the CSV" works okay enough, but if you task it with more than that, like see if a column of values in the CSV is entered correctly from the markdown, it breaks.

Or it gets stuck in a loop, and there's a very short point where it is faster to enter it by hand. Slightly ironic, though, that a language model doesn't do too well with natural language processing.

I'd certainly not trust it for anything important like a production database, but the csv/markdown thing isn't, and it's no big deal if it gets destroyed by the model/agent, so it's interesting to poke around with, and feel out the limitations, so you know its strengths and weaknesses.