You're not wrong, and to each his own. I'm a workflow and automation nerd, so it's down my alley personally.

These people are the ones causing the problems. They tell you they've got it all figured out make it sound great but they're the ones shitting out garbage faster than you can fix it.

I have hand-rolled agents/skills that run through a standardized process:

1. investigate
2. plan
3. execute
4. self-review
5. open PR

I leverage bash and python scripts whenever possible within harnesses, to try to avoid Claude injecting flair where steps and processes should be as repeatable as possible. I also write a lot of supplementary skills for it to use, with least-privilege scoping (important!), to give it access to do things like read work request tickets, run staging smoke tests, etc.

You can get to a point where you can basically say something like, "Investigate and plan SWE-1773" and it'll do all the shit, spit out a human-readable markdown plan complete with mermaid diagrams. You then okay to execute it, and at the end do an earnest PR review, and some manual tests.

Each brrr is gonna depend on how thorough you need to be, feature sizes, and if it shits the bed during planning. But honestly, the likelihood of your LLM shitting the bed is heavily dependent on how good of a harness engineer you want to become. If you shape and mold the environment and workflow like someone who say, works in developer tooling, then output quality can become surprisingly reliable.