Organizations deploying multi-agent AI systems are, in most cases, measuring the right things at the wrong level. An AI agent that resolves 65% of isolated software issues drops to 21% when tasks share dependencies across agent boundaries. Every individual agent passes every quality gate. The system still fails. Complexity science has a name for this pattern: when individually correct components produce collectively incorrect outcomes, the failure lives at the level of interactions, and standard component-level metrics are structurally blind to it. This talk draws on recent empirical findings and complexity science to show why quality metrics designed for individual artifacts and individual agents cannot capture the failure modes that multi-agent AI deployment produces in practice. I identify what organizations should monitor instead, at what scale of deployment the risk profile shifts non-linearly, and why current governance frameworks, including the EU AI Act, are not designed for systems whose failure modes live between agents rather than within them. The audience will leave with a concrete reframing of the governance question: from "is each agent safe?" to "what does this collection of agents produce that no agent was designed to produce?" That question now has both a theoretical foundation and the early outlines of a measurement approach.