The Context Access Divide: Interaction-Level Architecture as a Complementary Dimension of Agentic Inequality

Abstract: Sharp et al. (2025) introduce "agentic inequality" as a framework for analyzing disparities in access to AI agents across three dimensions: availability, quality, and quantity. These person- and organization-level dimensions characterize who can access agents and at what capability, but do not address a structurally important divide operating at a finer level: the individual interaction. Two users with nominally equivalent agent access may experience qualitatively different AI utility depending on whether the system can autonomously retrieve context from the user's knowledge corpus (Dynamic Context Retrieval) or requires the user to manually identify and attach relevant documents at each query (Manual Attachment). We term this the Context Access Divide (CAD). For knowledge-intensive workers whose intellectual capital spans tens of thousands of files, the CAD constitutes a qualitative threshold in AI usefulness: below it, the cognitive burden of context curation falls on the human, reproducing the inefficiencies AI is meant to eliminate. We propose contextuality -- the degree to which an AI system autonomously accesses a user's accumulated knowledge capital -- as a dimension of AI-mediated inequality that complements, but is not reducible to, the Sharp et al. framework. We formalize the CAD with a probabilistic model grounded in the fan effect literature in cognitive psychology, demonstrating that manual context attachment leads to a combinatorial collapse in task-success probability as corpus size and task conjunctivity grow, while dynamic retrieval architectures are structurally insulated from this collapse. We analyze the technical basis of this divide in the Model Context Protocol (MCP) and retrieval-augmented generation (RAG) architectures, and examine its implications for knowledge-work stratification and AI platform governance.

• The paper introduces contextuality as an interaction-level dimension of agentic inequality and models how manual context attachment can fail as knowledge corpora and task requirements grow.
• Its illustrative model shows that three-document task success falls from 0.51 at 50 documents to 0.0008 at 1,000, while Open DCRM remains near 0.86—about 5,300 times more likely to succeed.
• The analysis argues that Walled and Open DCRM architectures can reduce inequality by lowering context-curation burdens, while interoperability, portability, and better AI-use measures are needed to prevent platform lock-in and widening knowledge gaps.

Overview and contribution

This paper, by Masahiro Fujita (Kansai University), extends the "agentic inequality" framework of Sharp et al. (Sharp et al., 19 Oct 2025) by identifying a dimension of AI-mediated inequality that operates at a different level of analysis than that framework's three dimensions of availability, quality, and quantity. Sharp et al.'s dimensions are person- and organization-level variables describing who can access agents, at what capability, and in what numbers. The paper argues that a structurally significant divide is invisible at this level: whether an AI system can autonomously retrieve context from a user's knowledge corpus, or whether the user must manually identify and attach context at each query. The author terms this the Context Access Divide (CAD) and proposes contextuality—the degree to which an AI system autonomously accesses a user's accumulated knowledge capital—as a complementary inequality dimension.

The CAD is characterized as a macro-consequential micro-variable: it is a property of individual interaction architecture, but because it systematically advantages workers with large corpora and technical sophistication, its distributional consequences aggregate to the societal level. The relationship to Sharp et al. is one of cross-level articulation rather than parallel addition; contextuality names the mechanism by which interaction-architecture choices translate into the distributional outcomes that availability, quality, and quantity describe.

Three architectures of context supply

The paper distinguishes three architectural categories that cut across commercial providers:

Architecture	Context curation burden	Scope
Manual Attachment Model (MAM)	User recalls, locates, selects, attaches files per query	Only explicitly supplied context
Walled DCRM	System retrieves autonomously within provider ecosystem	Bounded by ecosystem (e.g., Gemini in Google Workspace; Copilot in Microsoft 365)
Open DCRM	System retrieves autonomously across ecosystem boundaries	Cross-provider, via open protocols such as MCP plus agentic RAG

The distinction between Walled and Open DCRM carries particular weight in the argument: both relieve the user of manual attachment within their scope, but only Open DCRM does so without imposing ecosystem dependency. Walled DCRM thus functions simultaneously as a capability improvement and as a lock-in mechanism—a point developed later into a governance critique.

The formal model and threshold dynamics

The analytical core is a probabilistic model of human recall limits applied to context curation, grounded in the fan effect literature [(2607.08495)'s citations to Anderson 1974; Anderson & Reder 1999]. For a task requiring $k$ conjunctively necessary documents from a corpus of size $N$, per-document recall probability $q(N)$ is modeled with a logistic decay motivated by fan effect findings, and MAM success probability is $q(N)^k$. Because $q(N) < 1$, success probability decays exponentially in $k$ while itself declining in $N$, producing a combinatorial collapse.

With illustrative parameters ($q_{\max}=0.95$, $q_{\min}=0.05$, $N_0=50$, $N$0), a three-document conjunctive task has MAM success probability of 0.51 at $N$1, falling to 0.0008 at $N$2 and approximately 0.0001 at $N$3. Under the same parameters, Open DCRM maintains success near 0.86 regardless of corpus size, implying that Open DCRM is roughly 5,300 times more likely to succeed on this task class than MAM at realistic corpus scales. Walled DCRM stabilizes at an intermediate plateau (~0.19 with $N$4), reflecting residual manual retrieval failure for out-of-ecosystem material. An appendix robustness check replaces the logistic form with an exponential decay motivated by semantic fluency research [Rohrer et al. 1995]; the collapse pattern is preserved and slightly steepened (advantage ratio 6,859×), suggesting the main specification may be conservative.

Two structural claims follow directly. First, the transitions among the three architectures are threshold effects, not smooth gradients: because knowledge-synthesis tasks exhibit conjunctive context dependency—a legal brief missing the two most on-point precedents may be affirmatively misleading—partial context yields qualitatively inferior output rather than proportionally partial utility. Second, the thresholds are nested: MAM → Walled DCRM → Open DCRM constitutes two qualitative discontinuities whose practical significance depends on where a user's knowledge capital resides. For users whose corpora are consolidated within a single ecosystem, the Walled-to-Open discontinuity may be negligible; the divide is most consequential for cross-ecosystem knowledge management, which the paper argues characterizes senior knowledge workers.

Positioning relative to existing frameworks

The CAD is partially correlated with each of Sharp et al.'s dimensions but reducible to none: a user may access a frontier model through a MAM-only interface, or configure DCRM atop a mid-tier model. Two workers with identical scores on availability, quality, and quantity may therefore face a qualitative discontinuity in AI usefulness invisible to the person-level framework. The paper also situates the CAD against third-generation digital divide scholarship and against Morisco's (Morisco, 25 Mar 2026) extension of the knowledge gap hypothesis: where Morisco locates new inequality in the evaluation of AI outputs, the CAD identifies a structural condition upstream—whether the system can access the relevant corpus at all.

Sub-components of contextuality include corpus scope, retrieval autonomy, corpus scale, temporal currency, and integration friction. The divide interacts with work type: it is most consequential for large-corpus, synthesis-intensive, non-linear knowledge work, and largely immaterial for bounded, self-contained tasks.

Social implications

The paper argues that DCRM access concentrates among already-advantaged actors through four mechanisms: organizational resources (Block reported engineers shipping over 40% more production code after internal MCP-connected agent deployment); technical capital (only 41% of senior technical leaders report any production MCP deployment per Stacklok's 2026 survey, implying far lower rates among non-technical workers); platform selection literacy; and corpus quality, which creates a compounding dynamic analogous to Merton's Matthew effect. A bidirectional feedback loop—DCRM accelerates knowledge accumulation, which increases DCRM benefit—has no natural equilibrium and may entrench stratification even if technical barriers decline.

A governance analysis follows. Providers have structural incentives to offer Walled DCRM because it combines retrieval benefits with lock-in, while Open DCRM generates no equivalent proprietary returns—implying systematic market under-provision of the architecture most beneficial to users with heterogeneous corpora. The Microsoft/OpenAI arrangement (over $13 billion invested; frontier models deployed primarily within walled integrations despite OpenAI's founding mission language) is presented as illustrative of this tension. Proposed policy instruments include interoperability mandates, data portability requirements preserving DCRM functionality across providers, and procurement standards favoring Open DCRM. The paper also notes that surveys treating "AI use" as binary will systematically underestimate variance in productivity effects—an implication for empirical methodology as much as policy.

Limitations

The paper is candid that its formal model is a theoretical illustration, not an empirical estimate. Parameters are chosen for plausibility, not fitted; the fan effect mechanism is validated only at small stimulus-set sizes (fans of 1–5 associates) and extrapolated several orders of magnitude beyond its measured range. The author argues this extrapolation cannot currently be tested within controlled-encoding paradigms, since feasible laboratory sessions cannot accommodate corpora of thousands of documents—and offers the suggestive, though not demonstrative, interpretation that this methodological inaccessibility itself corroborates the cognitive weight of the MAM task. The model assumes independence across document recall probabilities (positive dependence would attenuate the collapse), treats $N$5 as known in advance (real conditions are arguably worse), and omits false-positive attachments that dilute context windows. The recognition/cued-recall distinction is invoked to reconcile the model with Standing's massive-memory findings. More broadly, the paper is conceptual throughout: no ecological measurements of MAM-versus-DCRM productivity exist, the architecture boundaries are not sharp in practice, and the analysis presupposes a stability in the MAM/DCRM distinction that platform convergence may erode—even if transition-period inequalities persist after the gap closes.

Conclusion

The paper's central claim is that interaction-level architecture constitutes a genuine dimension of agentic inequality, not merely an ergonomic difference: for knowledge workers with large accumulated corpora, the CAD marks a qualitative threshold separating distinct categories of human–AI cognitive partnership. Its formal model, while uncalibrated, establishes that the combinatorial collapse under manual context attachment follows from multiplicative structure alone and is robust to functional-form choice. The most pressing open questions are empirical: direct estimation of recall probability $N$6 at professional corpus scales, measurement of productivity differentials between MAM and DCRM users in ecologically valid settings, and governance analysis sufficient to counteract the structural incentives favoring Walled over Open DCRM.