Agent-Generated User Interfaces: An Evidence-Classified Mapping of A2UI, GenUI, and Interoperable Agent-User Interaction

Abstract

Large language model agents create user interfaces for interaction with users at runtime. Evidence for these systems is spread across research papers, protocol specifications, engineering specifications, and other documents. This paper maps what is known about A2UIs, specifically A2UI and GenUI, and the strength of the evidence supporting each claim. A validated corpus of peer-reviewed papers, preprints, and grey literature was systematically mapped for each layer of technology, generation time, contribution to the A2UI, type of research performed, and class of evidence. The application dimension includes content automation and commerce. Technical and human evaluations focus mainly on code-generating interfaces. Declarative, catalog-based protocols have been evaluated mainly with regard to their specifications. There is thin evidence for incident response, multi-agent composition, mobile degradation, and cross-renderer interoperability. The mapping produced a repeatable coding scheme and a research agenda for semantic authorization, backend-to-UI state contracts, and governed commerce surfaces.