jetbrains junie is a lot better. but i still have yet to observe a series of performances from it better than 50% getting it right.

my advice about LLM coding agents is give them small specific tasks. they will cock it up more than a junior would, although they cock it up faster. after you watch it go wrong twice you just roll up your sleeves and start writing the thing it couldn't understand.

my biggest complaint is they are utterly incapable of architecture. my guess is that it would take mega-context (typical coding agents have around 20-50k context) to actually get that shit right. the agents can be improved more too, i think if the agent had a proper process for taking a whole mess and boiling it down over and over again until it comes up with an architecture diagram of the mess you want it to fix, but it probably would take at least all day on a moderately simple thing like a relay.