Self-Curated Agency: Digital Autonomy

• Self-curated agency is a paradigm where users, systems, or agents intentionally control their actions and representations through structured friction, traceability, and compositional tools.
• Design frameworks employ explicit decision points and provenance overlays, contrasting open algorithmic feeds with self-curated environments like Wikipedia and Are.na.
• Empirical studies, such as LIMI achieving 73.5% on AgencyBench with only 78 curated samples, demonstrate the performance benefits of targeted, high-quality curation.

Self-curated agency denotes the condition in which a user, system, or agent exercises intentional, situated, and structured authorship over its own actions, choices, or representations—distinct from purely externally sequenced or fully algorithmic behavior. In contemporary interface design, artificial intelligence, educational contexts, and computational modeling, self-curated agency is characterized by the explicit structuring of choice points (friction), fine-grained traceability of provenance, and compositional affordances that enable deliberate meaning-making and authorship. This paradigm stands in contrast to opaque, fully algorithm-mediated, or passively constrained environments, and is now technically grounded across interface analysis, formal models of agency, experimental evaluation, and multi-agent system design.

1. Conceptual Foundations: Hypertextual Friction and the Agency Model

Liu and Almeda introduce the construct of Hypertextual Friction as the organizing framework for self-curated agency in algorithmic environments (Liu et al., 31 Jul 2025). Hypertextual Friction operationalizes three core principles:

• Friction: Surfaces decision points in navigation or content creation, converting passive consumption into acts of authorship.
• Traceability: Exposes the lineage, edit history, source trails, and model influences underlying content, rendering assembly processes legible.
• Structure: Provides compositional tools (associative meshes, spatial canvases, editable blocks) for user-driven arrangement of knowledge.

Agency is modeled as a function of these properties: given friction $F$, traceability $T$, and structure $S$, user agency $A$ can be quantified via either a linear aggregation,

$A = \alpha F + \beta T + \gamma S,$

or a multiplicative form,

$A = \kappa (F \times T \times S),$

with $\alpha, \beta, \gamma, \kappa > 0$ weighting respective contributions. The linear model expresses additive contributions, while the product formulation asserts that the absence of any one dimension annihilates overall agency. Local friction at decision points calibrates the probability of transitions to reflect deliberate, non-automatic choice.

2. Comparative Paradigms: Hypertext vs. Algorithmic Systems

The comparison between hypertext systems (e.g., Wikipedia, Are.na) and algorithmic feed/recommendation systems (e.g., Instagram Explore, DALL·E) establishes the mechanistic basis of self-curation:

System	Friction	Traceability	Structure
Wikipedia	Link-based fork points; manual navigation	Citations, histories, visible edits	Mesh-like, multi-link topology
Instagram Explore	Infinite scroll, no explicit forks	Opaque ranking, no provenance	Linear, prescriptive feed
Are.na	Manual assembly of blocks, annotation prompts	Metadata per block, source URLs	Non-linear boards, link graphs
DALL·E	One-shot prompting, immediate outputs	No exposure of data or model flows	Flat, uneditable output arrays

Hypertext systems instantiate self-curated agency via visible choice architecture and provenance, supporting user-constructed knowledge trails and multi-path authorship. Algorithmic systems, by contrast, obscure process, flatten user participation, and preclude backtracking or remixing (Liu et al., 31 Jul 2025).

3. Design Commitments and Interaction Patterns

Designing for self-curated agency in contemporary systems entails embedding the following commitments (Liu et al., 31 Jul 2025):

• Deliberate Friction: Structure workflows such that users face presented forks, annotated options, and reflective pauses.
• Transparent Provenance: Attach source trails, history overlays, and model checkpoints to every item; allow in-line replay/remixing of content evolution.
• Compositional Structure: Provide non-linear canvases for open-ended mixing, clustering, and annotation; support thematic layering, fork-merge navigation, and authorship-graph attribution.
• Interpretative Over Prescription: Surface diverse options for user judgment, permit direct user tagging and annotation, and avoid reduction to a “single best” outcome.
• Curational–Generative Bridging: Require scaffolding inputs for AI-driven generation, offer undo/replay for generation steps, and ensure user arrangements become upstream influences for future outputs.

Illustrative patterns involve fork-and-merge navigation, provenance overlays, spatial associative canvases, and distributed authorship graphs.

4. Formal and Computational Models of Self-Curated Agency

Multiple formalisms operationalize self-curated agency at the system and agent level:

• Agency as Supervenient Downward Causation: Agency arises when a system’s macro (coarse-grained) variables, defined by the system’s own internal architecture, exert downward causation on micro-level dynamics, and the system maintains a predictive gap between its own anticipations and realizations. The agency index is

$$A = \mathbb{E}_t[\Delta^{in}_t] \times \mathbf{1}\{\Delta^{ex} > 0\},$$

where $\Delta^{in}_t$ is internal regulation error, and $\Delta^{ex}$ is external irreducibility, marking genuinely self-generated teleological constraints (Horibe et al., 7 Dec 2025).

• Self-Governing Dynamical Laws: In dual-level models, the agent possesses not only a physically-closed base-level dynamic but also an independent supervenience-level law $P$ determining composite behavioral sequences $c_t$, so that changes at the agent level can causally influence base dynamics via feedback error (Ohmura et al., 6 Jan 2026). The locus of moral and functional responsibility is thus the agent’s internal “law” for updating $c_t$.
• Collective and Authorial Dimensions in Learning: “Self-curated agency” in educational AI is expressed as a weighted sum of authorial agency, dynamically emergent agency, agentic engagement, and mini-c creativity:

$$\text{SCA} = \lambda_1 A_{auth} + \lambda_2 A_{dyn} + \lambda_3 AE + \lambda_4 C_{mini},$$

where each component is formally modeled to capture dialogic authorship, co-construction, multi-level engagement, and personal meaning (Dai, 8 Dec 2025).

5. Empirical Evidence, Benchmarks, and Practical Engineering

Empirical approaches to cultivating self-curated agency include:

• Algorithmic Curation and Agency-Scaffolded Models: LIMI demonstrates that strategic curation—not sheer data volume—yields superior emergent agency in LLMs. With only 78 curated demonstration trajectories, LIMI achieves 73.5% on AgencyBench, surpassing models trained on over 10,000 samples (47.8%) (Xiao et al., 22 Sep 2025). This establishes the Agency Efficiency Principle: agency is maximized via focused, high-quality demonstrations.
• Dynamic Alignment with Self-Curated Policy: The Dynamic Alignment framework for Collective Agency aligns LLMs in a fully self-supervised regime. Automated prompt generation, iterative self-critique, and self-rewarding via GRPO enable the model to refine itself toward the open-ended ideal of sustaining and empowering agentic capacities across knowledge, power, vitality, and benevolence (Anantaprayoon et al., 5 Dec 2025).
• Design Evaluation in Human-Computer Interaction: Empirical studies in media interfaces show that self-curated agency correlates with explicit planning (playlist assembly, search), graduated control (user-determined recommendation exposure), and informed choice (surfacing metadata) (Lukoff et al., 2021). Conversely, seamless algorithmic push (autoplay, unmodulated recommendations) undermines user agency.

6. Challenges and Theoretical Tensions

Self-curated agency presents a fundamental tension within data-driven systems. Granting users or agents maximal agency—via self-selection or layering of autonomous laws—generally undermines the identifiability of causal effects in observational research, as demonstrated by systematic sign-mismatches between self-curated observational estimates and randomized ground-truth in online platform experiments (Milli et al., 2021).

The “Catch-22” is that fully respecting self-curated agency (embracing user autonomy and consent) makes statistical inference from unmodeled preferences intractable, while the imposition of constraints needed for identifiability negates authentic agency.

7. Toward Autonomous, Creative, and Self-Sustaining Agency

Constructs from autocatalytic network theory (RAF theory) formalize the emergence and persistence of self-curated agency in AI. An autonomous agent achieves self-curation when it possesses a reflexively autocatalytic, food-generated (RAF) core: a set of operations both catalyzed and sustained internally, capable of absorbing novelty through integration of external or internally generated tasks. Phase-transition theorems specify the critical threshold for emergence of RAFs, pinpointing when a system transits from externally driven to genuinely self-maintaining (Gabora et al., 2024).

Architectural blueprints for agentic AI require:

• Permanent memory for primitives,
• A persistent RAF kernel,
• Novelty evaluators and integration mechanisms,
• Self-monitoring to maintain closure and catalytic support.

Persistent, self-curated agency is marked by the system’s ability to extend its internal core in response to challenge, thus solidifying self-identity and supporting creative transformation.

---

References:

(Liu et al., 31 Jul 2025, Anantaprayoon et al., 5 Dec 2025, Lukoff et al., 2021, Dai, 8 Dec 2025, Ohmura et al., 6 Jan 2026, Milli et al., 2021, Horibe et al., 7 Dec 2025, Sharma et al., 2023, Xiao et al., 22 Sep 2025, Gabora et al., 2024)