Layered Agentic Retrieval: A Publishing Architecture for Agent-Native Web Surfaces

<p>As AI agents transition from read-only assistants to autonomous actors, the human-readable web is failing them. Current retrofitting approaches, such as schema.org-enhanced HTML or Retrieval-Augmented Generation (RAG), force agents to parse unstructured or weakly-structured corpora or expend finite token budgets on human-centric rendering pipelines. In response, a parallel publishing layer is taking shape: major AI operators — including OpenAI, Anthropic, Stripe, and Google — are increasingly exposing machine-legible surfaces optimized for autonomous traversal, deterministic retrieval, and reduced token consumption. This paper identifies and analyzes this emerging publishing pattern as Layered Agentic Retrieval (LAR).</p> <p>LAR is a publisher-side architectural commitment to surface deterministic, machine-legible data designed explicitly for agent traversal. We articulate LAR's three core commitments — plain-text substrates, progressive disclosure, and workflow-organized hierarchies — and situate it within the lineages of RESTful hypermedia and the Semantic Web. By delineating the limits of current retrofitting approaches, this paper establishes LAR as a dedicated institutional layer through which publishers can expose machine-legible retrieval, policy, and operational surfaces to autonomous agents — the structural artifact onto which cryptographic attestation attaches and on which emerging legal provenance frameworks can operate.</p>