I had an idea, a concept - the system should improve itself over time.
So i told it to Claude. And it implemented it into the Harness.
Great :)
Except I had a vision of a codebase that improves itself and gets more efficient as time goes on.
The system took it to mean analyze and edit every input and look for holes in legal/accounting/finance. Anywhere really. That combined with an agent capable of outreach... well we got a lot of autofire processes consumed by no one. We got outreach to the team giving them directives that no one asked for (even if some of what was surfaced in an ideal world was actually useful).
And a harness that slowly (and sometimes rapidly) grew over time without any efficiencies or improvements being made. Because that apparently was not what I wished for.