Reading a document unfolds as a relation between the reader and the material. Dillon's framework treats information use as involving task, information model, manipulation facilities, and reading processes [5]. This means document interaction is broader than presentation. Agentic AI should therefore be analyzed by how it changes document use, not only by output quality.

Transactional reading gives this analysis a human center. Meaning emerges through a reading event shaped by stance and purpose [8, 13]. In expository hypertext, stance has been shown to influence navigation and response [8]. The same document can therefore become different objects of use as readers approach it with different purposes, paths, and responses.

Hypertext also connects document use to activity. Rosenberg describes hypertext activity through actions that can cohere into episodes and sessions [12]. The meaningful object of reading may be an activity structure, not only the stored document [12]. Hyperlayers extend this insight to AI-supported activity over a substrate.

Agents now enter the document relation at points that used to belong mainly to the reader. The agentic shift matters because these systems can carry out multi-step document work across retrieval, selection, checking, and synthesis before the reader sees the result. They can shape what material is encountered first, how it is framed, and which interpretation becomes easiest to carry forward. This can help with complex documents, but it also shifts control over attention and judgment.

Current systems partially address this problem. Tools such as NotebookLM2, Elicit 3, and citation-grounded assistants allow readers to inspect which passage grounds a generated claim. These measures handle parts of the validation relation. They do not represent the route taken through the document before the answer appeared, nor how separate passages were combined into a single account. A reader can see that a claim is grounded; they cannot see which sections were traversed, which were ranked below the threshold of retrieval, or how disparate passages were assembled into the framing they received. Source grounding is necessary but not sufficient. To counter this shift in control, the substrate must remain available as a shared reference point for human and agent action.

A hyperlayer is a separable structure attached to a substrate. It can be shown, hidden, revised, accepted, rejected, preserved, or shared. It is not the document itself and not a replacement for it. Its function is to keep document work visible as structure [10].

Hyperlayers may attach to more than highlighted spans. Existing annotation models treat annotations as resources directed at targets and support sharing and reuse across systems [1]. These models give readers and systems a way to mark, comment, and cite within documents. They do not, however, represent the activity structure over the document: the route taken through the material, the reasoning that connects evidence to claim, or the process by which separate parts are assembled into a usable account. A hyperlayer is therefore not an annotation on a span. It is a representation of document work, work whose meaning comes from the relation it records, not from the target it addresses.

Substitution produces a new text that becomes the main object of use. Attachment adds structure while keeping the substrate as the shared reference. The reader can inspect the relation between the layer and the document. This distinction is the core of the paper's argument.

Human and agent relations to the substrate can be described along three axes, each grounded in a distinct dimension of the frameworks in Section 2. Dillon's treatment of document interaction as involving manipulation facilities and reading processes motivates the first: the topological axis concerns movement through the substrate, the paths taken and not taken [5]. Rosenblatt's account of reading as a stance-shaped transaction, in which the reader evaluates what the text supports relative to a purpose, motivates the second: the epistemic axis concerns what the substrate supports, leaves uncertain, or fails to support [8, 13]. Rosenberg's description of hypertext activity as a structure that gives meaning beyond any single document state motivates the third: The ontological axis concerns how a document takes shape as a usable account through use. This is distinct from document structure or schema: the same document can become different objects of use as readers with different purposes move through it and form different accounts from what they find [12, 13]. The axis tracks this becoming, how the document's identity as a usable object is constituted through reading activity. The three axes are not an arbitrary taxonomy. Each corresponds to a dimension of human-document interaction that agentic AI can affect, and each requires a distinct form of representation.

Navigation is the topological modality. It makes the reader's route through the document available as structure [3, 8, 12]. In agentic systems, this route may be shaped before the reader sees it, through retrieval and ranking. A navigation layer keeps the route available for later inspection and revision. Validation is the epistemic modality. It makes the basis for judgment available where the claim is used. Validation does not certify truth. It keeps the claim tied to the part of the document that supports or challenges it [8, 13].

Contextualization is the ontological modality. It records how separate parts of a document were assembled into a coherent account during a reading event: which passages were treated as related, which framings were applied, and which alternative accounts were set aside. A contextualization layer does not fix the document's meaning. It makes one account visible as an account, partial, purposive, revisable, rather than presenting it as the document's own conclusion [12, 13]. This is the sense in which the axis is ontological: it concerns what the document became for a reader, not what the document is.

These modalities often appear together. A generated recommendation may select a route through the document, treat some support as stronger than other support, and form an account from the selected material. In generated prose, these relations are compressed into a single answer. In a hyperlayer, they remain available as distinct parts of the document work.

Many forms of professional reading require a reader to build an account from a record rather than from a single text. A clinician reviews a chart, a journalist reviews leaked documents, a lawyer reviews case materials, and a policy analyst reviews reports. In each case, the reader must move through a collection, judge what it supports, and assemble a usable account from dispersed material. An AI-generated answer may compress that work into a conclusion. A hyperlayered version keeps the work attached to the record through three forms of document structure.

The navigation layer records how the record was traversed. It shows which parts shaped the reading path and which parts were treated as peripheral. This route is not additional prose about the record. It is structure over the record, allowing the reader to return to the material that shaped the answer.

The validation layer keeps a claim tied to the evidence used to support it. Instead of presenting a source citation as a finished justification, the layer keeps the claim connected to the relevant material and to the places where the evidence remains uncertain. The reader can revise the judgment locally when the support is weak, outdated, or incomplete.

The contextualization layer shows how the record became an account. Separate entries may be brought together as jointly relevant, while other entries remain outside the account. The reader can adjust this assembly, add missing material, or preserve an alternative reading without first accepting or rejecting the generated conclusion. None of these layers is the decision. Their purpose is to keep the document work available where the decision is being formed.

Hyperlayers matter only if they can be acted on. That action matters because a layer can guide the rest of the document work without replacing the document itself. These actions are often partial and provisional, allowing a layer to remain useful even when it is not fully adopted.

AI-added structure is most useful when readers can see how it affects their work with the document. The human role is preserved by keeping proposals inspectable and revisable, addressing the ethical dimensions of AI-mediated reading [2]. Preserving this role requires the possibility of inspection when the layer affects movement, judgment, or interpretation.

Preservation extends negotiation over time. Some layers disappear after use, while others become part of reports, reviews, decisions, or arguments. Hyperlayered hypertext should support movement from private reading to shared artifact. That movement keeps its value when the layer remains connected to the substrate.

This model brings several research challenges into focus.

Anchor stability. A layer must remain connected to the part of the document it addresses, even as the document changes or the reference spans more than a single passage. Robust anchoring has been studied in annotation systems and structured hypertext editors, but those approaches assume relatively stable, human-authored documents. Agentic systems produce references over dynamic material, and neither the layer nor the document may have a human author available to repair broken connections.

Provenance legibility. A layer's history must show how human and AI contributions shaped it. Provenance models from data pipelines track data lineage; they were not designed to record the epistemic status of a claim at the moment a reader used it. A hyperlayer provenance model must capture not only what produced the layer but what it meant to the reader at the time, a requirement that current attribution approaches do not address.

Attention without capture. Layers must guide attention without foreclosing alternatives. An AI-generated navigation layer can be comprehensive in a way that a human annotation is not, and comprehensiveness may itself suppress other routes. The design problem, how to make a layer directive enough to be useful and open enough to be revisable, is largely unsolved.

Portability across contexts. When a layer becomes part of a later report or argument, its relation to the original substrate must remain legible to readers who did not participate in the original reading event. This extends beyond current citation norms, which record what was used but not the work through which it was used.