Agent Behavioral Contracts II: agents co-fail on 90% of missions, so composing reliability bounds under independence is wrong

arXiv·high signalXBlueskyLinkedInCopy link

arXiv 2608.12895 (2026-08-13) reports a preregistered evaluation over 18,000 missions and finds that when either of two agents fails, both fail on 90.0% of those missions (log OR 6.66, 95% CI [6.38, 7.00]). That positive dependence directly invalidates the independence assumption behind the usual compositional reliability math, which means multiplying per-agent success rates systematically overstates a pipeline's reliability. For anyone sizing redundancy into a multi-agent system, adding a second agent buys far less than the arithmetic suggests.