Agent Behavioral Contracts II: Certifying Compositional Reliability Without Assuming Independence

Abstract

The Agent Behavioral Contracts (ABC) framework certifies a composed AI-agent pipeline's reliability by a compositional lower bound valid under five conditions, the last of which (C5) is independence of stage failures. This paper asks what a compositional certificate can promise when C5 is false - as it routinely is, because multi-agent pipelines are commonly built from several instances of the same frontier model. On real retail and financial agent tasks scored by deterministic gold code, we measure large cross-agent failure overlap (74-90% of failures shared; a capability-diverse cross-architecture pair shows perfectly nested failures, Fisher exact p = 3e-43). We show two natural certificates are unsound: the independence product is anti-conservative for redundancy, and a fitted Gaussian-copula floor is not a finite-sample guarantee - its coverage of the true reliability collapses to zero as data accrue (a partial-identification obstruction, reached concurrently for parallel ensembles by Chen 2026). We give a sound tiered certificate: an exact Clopper-Pearson bound on the directly observed all-success event, and a copula-agnostic linear program over the identification set that tightens monotonically as more co-execution moments are supplied. With the v1 patent claim withdrawn, the full ABC framework and every formula are disclosed. This work is a contribution to AI reliability engineering. Companion to arXiv:2602.22302.