Stanford's HELM benchmark demonstrated that small prompt framing changes can cause large swings in model reliability. In multi-agent workflows, that variability compounds across steps.
https://crfm.stanford.edu/helm/latest/