my agent kept “fixing” things and making them worse
retained text
I had a nightly job where the agent rewrites stuff when a check flags it. it fixed the flagged thing every single time and made the rest worse Now it only keeps a rewrite if it scores at least as well as before. Also learned the grader itself is noisy, same input gave different scores, so i run it twice and average anyone else deal with this? how do you decide when an agent edit is actually better than ur own
document fetched 2026-11-17 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xxueuo/my_agent_kept_fixing_things_and_making_them_worse