"You're absolutely right, I should have..." (fourth day running)
Wrote a #blog about making my AI agent fix its own harness instead. Every two hours an outside judge reads its transcript, each mistake becomes a hook or a test, and later runs watch for repeats. Two days in, 31 fixes already failed.
https://cloudrumble.net/blog/recursive-harness-self-improvement
cloudrumble.netYou're Absolutely Right: Now Fix Your Harness | Cloud RumbleThe model is fixed and the harness is mutable. How an observer, an independent judge and an improving agent turn my Claude Code assistant's mistakes into hooks and tools with regression tests, and what a quiet day after a change does and does not show.