14 failure modes in 3 categories across 1,600+ annotated execution traces from 7 multi-agent frameworks; a taxonomy, not an accuracy benchmark.
discriminates
[error-decorrelation][producer-agreement]
saturation
n-a
patternLabFit
Instruments what passRate measures only implicitly; its inter-annotator Cohen's kappa of 0.88 is the exemplar for validating human annotation of collaboration quality. The inter-agent misalignment category maps to producerAgreement and cohensKappa.
