It's like an over-eager junior engineer that will implement whatever pattern they read about last week. It just doesn't feel trained on high-quality enterprise, battle-hardened codebases.
Even Opus 5.5 on High thinking still littered a root directory of a JS project with source files without thinking to put it in a 'src' directory. It had to be told to use a testing framework. Had to be told to also test with Firefox (the project already had established Chrome /and/ Firefox targets).
I have to continuously burn tokens and effort correcting it.
I imagine it performs a bit better with a huge corpus of architecture and style guides. But I did expect to get some that for free with the "leading" model.
Same with that situation. One shot code from prompt is still not great. Maybe if you had a prompt with highly detailed requirements for structure and code formatting it would be ok.
I generally hate the look of AI gen code, so I have some pretty detailed code formatting rules for even small edits. And the AIs still ignore them some of the time.
I don't know what I'm getting at here; other than. It is acting pretty human like at this point in the code delivery phase. Human as in asking someone to do your work for you and you not always getting back what you want; as opposed to being good enough to replace your actual work (assuming you are good at the task you are asking from the Agent -- does work pretty fast though. which makes it a trade off worth considering; but then your time is in picking through the code to confirm it is correct).