This paper introduces fragmented intention - a structural condition in distributed AI agent ecosystems where no single model invocation contains sufficient information to infer the global objective of a coordinated action sequence. We prove that under such conditions, model-level safety mechanisms become structurally insufficient (Detection Collapse Theorem), and propose the Autonomous Systems Coordination Layer (ASCL) as a coordination framework binding persistent identity, behavioral trust, economic accountability, and capability-scoped interaction. Includes formal proofs, Monte Carlo simulation (10,000 trials), and a normative specification.