Paper T3 — Decomposing Agency, Isolating Answerability: Cultivating What Cannot Be Delegated in AI-Assisted Learning

Preprint version: 2
Canonical public preprint DOI: https://doi.org/10.35542/osf.io/hvbfe_v2

Kengo Tomita
Institute of Technology, Shimizu Corporation, Tokyo, Japan

Abstract

When AI can plan, draft, and revise, what remains for the learner? This article argues that the question cannot be answered while agency is treated as monolithic, and decomposes learner agency by delegability — jointly with its educational companions, trainability and measurability — rather than by psychological function. Three analytic dimensions are provisionally distinguished: direction (evaluative orientation), drive in two forms (available activation and motivational drive), and mode (functional pattern of epistemic action). Mapped this way, most components admit support, training, or conversion, specifying where distribution accounts of human–AI agency apply. The decomposition also exposes a question that no capacity-based analysis can answer: who must answer for the judgments made with these capacities. That question isolates answerability — the non-transferable standing to have to answer, to those entitled to ask, for one's judgments — grounded second-personally by extending Darwall's address structure from conduct to the warrant of judgment. A design conjecture proposes that embodied encounter may support the crystallization of direction and the conversion of available activation into motivational drive, addressing formation within the capacity layer. A two-layer prescription follows: an analytic compass, encounter engineering, and AI guardrails for what can be distributed; protected second-personal exchange for what cannot. AI can help learners rehearse the competence of answering; the final normative addressee remains a legitimate human addressee — an educator, peer, professional body, or represented community.

1. Introduction

Generative AI has made learner agency the contested center of educational technology. When systems plan, draft, revise, and explain on request, the question of what remains for the learner to originate stops being rhetorical: capacities that instruction once cultivated because no tool could supply them must now be justified anew, and agency has become the word under which that justification is fought. Institutional responses so far oscillate between prohibition and laissez-faire — banning the tools to protect the capacities, or admitting the tools and hoping the capacities survive. Both responses treat agency as a single quantity that AI either threatens or does not. That premise is what this article examines.

Three streams converge on the concept without resolving it, because each decomposes agency for its own question. The first asks which systems qualify as agents at all. Dung (2025) shows that this field already treats agency as graded and multidimensional, profiling systems along goal-directedness, autonomy, efficacy, planning, and intentionality. But each dimension is a capacity the system possesses — even intentionality, the most demanding, asks whether a system owns and can reflectively revise its reasons, not whether anyone must answer to another for them. The vocabulary of answerability never appears in the account, and moral agency is explicitly set aside.

The second stream relocates agency from individuals to configurations. Cukurova (2026), introducing a British Journal of Educational Technology special issue on agency and self-directed learning with generative AI, argues that agency in AI-rich settings is best read as a system property — distributed and re-arranged across layered human–AI couplings, from the transactional to the synergistic — and closes with a demand this article accepts: agency must remain measurable and protected, not a residual aspiration. Notably, at the top of that account sits the learner's capacity to author meaning in ways that remain recognizably one's own, and a learning that is relationally responsible. The account is built to keep agency analyzable as capability migrates into the system, and it treats even these as measurable system properties. What it does not ask is whether everything gathered under agency migrates — and, as Section 4 argues, what appears at the summit of that account (whose judgment it remains, who must answer for it) is what a finer distribution analysis isolates rather than distributes.

The third stream is education's own: agency as an individual capacity to be fostered — resolved by psychological function into intentionality, forethought, self-reactiveness, and self-reflectiveness (Bandura, 2001); by temporal orientation into iterational, projective, and practical-evaluative dimensions (Emirbayer & Mische, 1998); with self-regulated learning supplying its most trainable core (Zimmerman, 2000). These accounts were built for human learners among human teachers. They say what agency is made of, not what may be handed over.

The streams are usually allowed to blur, because agency travels between them unmarked. A recent review consolidates the social-science treatments — Cathcart et al. (2026) sort fifty-one studies into psychological, sociological, and emergent (ecological) conceptions and argue for the emergent approach; Cukurova's distribution reading leans the same way. Those typologies map the terrain agency theory already occupies; none was built for the question AI now forces. This article does not add a fourth conception alongside them but cuts on a different axis — delegability — to answer that question. The streams should be held apart: the first concerns the attribution of agency to machines; the second its distribution across hybrid systems; this article concerns learner agency — the thing education exists to develop. Held apart, the gap becomes visible. None of the three asks the question AI forces on education: of everything gathered under learner agency, which components can be delegated, trained, or supported — and what, if anything, cannot? The same gap shows on the relational side. That education is relational is well established, from dialogic and care-ethical traditions onward; online participation modes alter opportunities for interaction, and AI-mediated assessment may affect trust — patterns documented largely through psychological and affective vocabulary (presence, belonging, connectedness, trust) — dimensions whose importance these studies document (Luo, 2025; Kortemeyer et al., 2023) — rather than as a normative relation in which one must answer for one's judgments. What is under-theorized is not that something relational matters, but the normative structure of the relation itself. Dialogic and care-ethical traditions identify the educational importance of relation; the narrower claim here is that generative AI makes newly visible the delegability of one specific relation — the one in which a learner has to stand behind a judgment as theirs (the warrant-bearing relation).

That question needs its own axis, and supplying it is this article's contribution: we decompose learner agency by delegability rather than psychological function, show that most components are trainable, supportable, or distributable across human–AI systems, and show that the decomposition exposes a question no capacity inventory can answer — who must answer for the judgments made — whose answer is answerability, a relation rather than a component, yielding a two-layer educational prescription. The move extends a line of work on AI-assisted knowledge work in which externalization (the work of turning judgment into artifacts) was first separated from specification (the judgment of what should be made) (Tomita, 2026a), and specification was then resolved into measurable and non-measurable components (Tomita, 2026b): the operator is constant; the contested concept changes. In that line's original analysis, learner agency was held constant within the human expert condition; the present article opens that background variable. Section 2 performs the decomposition and the sort. Section 3 proposes a design conjecture for the component that resists installation. Section 4 establishes what the decomposition isolates, and grounds it second-personally. Section 5 derives a prescription across two pedagogically concurrent layers, stating in each case what is deliberately left to those who must answer for it.

Relation to adjacent research. Four research traditions border this territory, and the axis of this article differs from each. Self-determination theory explains how external goals become autonomously endorsed through supports for autonomy, competence, and relatedness (Ryan & Deci, 2000); it concerns the quality of motivation, not who must answer for a judgment. The four-phase model of interest development traces how situational interest deepens into individual interest (Hidi & Renninger, 2006); it describes the formation of direction, and this article's encounter conjecture is a design-oriented cousin of that account rather than a rival. Self-regulated learning and its socially shared extension model how learners plan, monitor, and evaluate, alone and jointly (Zimmerman, 2002; Järvelä & Hadwin, 2013); these are capacities and practices — in the present vocabulary, largely supportable and trainable components. Epistemic agency in knowledge building assigns learners collective cognitive responsibility for advancing communal knowledge (Scardamalia, 2002); it is the nearest neighbor, yet its unit is the community's epistemic progress, whereas answerability is a dyadic normative standing: a particular learner answering to a particular questioner for a particular judgment. These traditions do not foreground the specific distinction developed here — between distributed performance and the identifiable bearer of a judgment's warrant. The contribution is therefore not the discovery that educational relations matter, but the articulation of a delegability distinction within them.

Two layers, announced. The argument proceeds on two layers that must not be conflated: a capacity layer, where components can be supported, trained, or transformed under specified conditions — and where AI can legitimately participate — and a relational layer, where a learner stands answerable to particular others. The decomposition in §2 works entirely on the first layer; §4 shows what it exposes on the second.

2. Decomposing learner agency by delegability

2.1 The choice of axis

Decompositions serve questions. When Bandura (2001) resolves agency into intentionality, forethought, self-reactiveness, and self-reflectiveness, the axis is psychological function: the question is what agency consists of as a human achievement. When Emirbayer and Mische (1998) resolve it into iterational, projective, and practical-evaluative dimensions, the axis is temporal orientation: the question is how agentic action relates past, future, and present. Both decompositions answer their questions well. Neither was built for the question generative AI now forces on education: of everything gathered under "learner agency," what can be handed over — to an AI system, to a human–AI configuration, to anyone other than the learner — and what cannot?

This paper cuts along that axis. We decompose learner agency by delegability — jointly with its educational companions, trainability and measurability — rather than by psychological function. The same material, cut along a different plane, yields different cross-sections; overlap between the components below and established constructs of motivation, interest, or self-regulation is therefore expected, and is not the contribution. The contribution is what the new plane separates that the old planes keep together: functions that can be supported or partially offloaded, a component whose endorsement is non-substitutable, and a relational question that no component inventory can answer.

2.2 Three components

Cut this way, learner agency can be provisionally distinguished into three analytic dimensions — a working schema warranted by intervention divergence rather than claimed as a psychologically exhaustive taxonomy. Direction is what the learner is oriented toward. It exists in two states: a candidate direction — an orientation entertained but not yet owned — and an endorsed direction, the evaluative commitment that makes some problems theirs rather than merely assigned; endorsement may be co-authored in dialogue rather than reached alone. The endorsed state is the educational counterpart of the Sollen-type — that is, value-laden, evaluative — content that prior work located at the non-delegable core of human specification (Tomita, 2026b). Both distinctions are defined within this article as they are used, so no step of the argument presupposes either paper. Magnitude is the drive behind the orientation: how much force the learner brings. Mode denotes the repertoire of epistemic actions a learner currently tends to deploy — whether they typically create, connect, critique, or maintain — rather than a fixed learner type; the examples are illustrative, non-exhaustive, and task-sensitive. The learner who sharpens any draft put in front of them but originates none is not low in agency; they are operating in a single mode. The triad is deliberately coarse. These are sorting bins for a delegability analysis, not a new psychology of agency; each could be subdivided further, and the subdivisions would matter for other questions. For this question, three suffice, because the components already behave differently under the only test that matters here: what happens when one tries to hand each of them over. The warrant for the triad is not psychological exhaustiveness but intervention divergence: each category directs education toward a different lever under AI assistance.

2.3 Two forms of magnitude

Magnitude is not a single quantity. Educators will recognize the state: the learner with substantial energy and no settled object — I want to do something; I do not yet know what. We call this available activation: the learner's current, context-sensitive capacity to mobilize effort, varying with health, affect, workload, accessibility, and circumstance — not treated here as a fixed trait, nor yet anchored to any particular orientation. If drive were simply a function of direction, this state could not exist. It does, and two further observations confirm the split it implies: learners pursuing the same direction differ persistently in the force they bring to it, and learners who change direction often carry substantial energy into the new field. We therefore distinguish available activation from motivational drive: propulsion that exists only as anchored to a specific direction. The two forms respond to different levers. Available activation can be protected, provisioned, and depleted, but not lectured into existence. Motivational drive is derivative: it appears when a direction captures available activation. How that capture happens is the subject of Section 3. What matters for the sorting is that the two forms separate cleanly under delegability.

2.4 The sort

Delegability is applied here jointly with its educational companions — trainability (can deliberate practice develop it?) and measurability (can its current state be observed?). These are distinct operations, not values on a single scale; the sort records, for each component, which operations it admits. Mapped this way, the components sort cleanly.

Available activation is supportable and protectable. Institutions shape it through workload design, wellbeing and accessibility provision, recovery time, and the removal of depleting friction — the externalization-cost reduction that AI now performs at scale belongs here, on the support side (Tomita, 2026a). Institutions cannot reliably manufacture available activation by exhortation; the defensible levers are protection and provisioning — reducing depleting friction, designing workload and recovery — not production on demand. Nothing here licenses treating low activation as a fixed attribute of the learner, and this analysis is not proposed as a basis for admissions, selection, or deficit classification.

Mode is trainable. A learner whose drive habitually critiques can learn to create; repertoires widen with practice and narrow with disuse. AI systems can scaffold the weaker modes directly — drafting so the critic has an object, structuring so the connector has materials — which makes mode supportable as well as trainable, and the least contested entry in the sort.

Motivational drive is convertible. It is neither trained directly nor installed from outside; it is converted from available activation when a direction takes hold. The conditions under which this conversion may occur are addressed by the design conjecture in Section 3.

Direction alone resists. Here the sort requires precision, because what resists is not the supply of candidate directions. Proposals are cheap: a teacher can suggest ten, an AI can generate fifty on request, and both have a legitimate role in widening what a learner encounters. What cannot be handed over is ownership. A proposed direction becomes the learner's own — or fails to — through a process the proposer does not control, and the difference between an installed direction and an owned one is behaviorally visible: compliance tracks the supervisor's attention; commitment does not. Self-determination theory documented this boundary a generation ago from a different angle: externally proposed goals become integrated, rather than merely complied with, only under conditions that are relational as much as informational (Ryan & Deci, 2000) — a finding whose second-personal undertone Section 5 returns to. Direction is not installed. It crystallizes as the learner's own, or it remains someone else's.

Table 1 summarizes the sort.

Table 1. Learner agency under the delegability axis.

Note. Answerability is omitted because it is a relational status rather than a learner component. The delegability question nevertheless applies: its institutional allocation may change, but answering for a judgment cannot be performed by a substitute while the judgment remains the learner's own. Section 4 establishes this standing.

2.5 What the sort shows about distribution

Set this result against the distribution account of agency. Reading agency as a property of human–AI systems rather than of individuals (Cukurova, 2026) is right — indispensably right — about most of the list above: planning and execution scaffolds, support structures for available activation, mode-widening tools, even the maintenance architecture around an existing motivational drive all genuinely distribute across hybrid configurations, and the sort specifies exactly where that reading holds. What the sort adds is the location of its silence. A hybrid system can host proposals for direction in any number; what it cannot do is supply an individual learner's endorsement while the resulting judgment is still presented as that learner's own. A hybrid system may participate in forming and sustaining a shared direction; what cannot be substituted is each learner's endorsement of the commitment they claim as their own. Distribution accounts are right about what distributes — and silent about what does not. The present article is therefore not a rejection of distributed agency but a component-level restriction on what distribution can legitimately explain. And the sorted residue hands the next section its question: if direction cannot be installed, where does it come from?

3. A design conjecture: encounter, crystallization, conversion

If direction cannot be installed, where does it come from? We propose a design conjecture in three movements: embodied encounter with a problem-holder may contribute to the crystallization of evaluative commitment, and an endorsed direction may then anchor available activation as motivational drive. The model is deliberately spare — each movement names a condition, not a guarantee.

The metaphor is chosen for its precision, not its color. Crystallization needs a supersaturated solution and a nucleation site: the available activation of Section 2 is the supersaturation — energy without an object — and the embodied encounter supplies the site. Read this way, the model makes intelligible the two failure modes educators already know. Exhortation without encounter — telling an unengaged learner to find their passion — addresses a solution that has nowhere to crystallize: there is energy, but no site at which a commitment could form. Encounter without protected follow-through — a moving field placement absorbed back into a packed schedule of unrelated demands — dissolves what had begun to form. Neither failure is a motivation deficit, and neither is fixed by more exhortation.

Three heterogeneous illustrative sources display sequences compatible with this conjecture. The innovation cases assembled by Nonaka and Takeuchi (1995) show the pattern repeatedly: developers' commitments formed at sites of direct contact with users and their troubles, not in planning meetings, and the organizational machinery that mattered was the machinery that put people in front of problems and then protected what crystallized there. The publicly documented winner profiles of a recent global AI-assisted development competition show the same pattern from the opposite end of the tooling era: among the top entrants, a lawyer automating building permits, a cardiologist building post-visit care, a musician, a road engineer — practitioners whose commitments predated their tools, for whom AI removed the externalization barrier and changed nothing about where the direction had come from (in a hackathon hosted by an AI developer; Anthropic, 2026). And the single-case trajectory documented in prior work (Tomita, 2026a) shows the sequence at fine grain: a domain commitment formed in the researcher's own usage context steered the tool, not the reverse. The three sources share no method, population, or era; what they share is the shape of the sequence. Jointly, they complicate the two default stories — that direction is installed by instruction, or that it is fixed temperament — since in each case it postdates an encounter and predates the tooling. These sources are illustrative and hypothesis-generating rather than confirmatory — the third, in particular, is the author's own prior work and cannot count as independent convergence.

We present this as a design conjecture motivated by illustrative, hypothesis-generating observations — not a demonstrated mechanism. No controlled study here isolates encounter as the cause of crystallization, and the sources above were not collected to test the model they now illustrate. The model earns its place by what it organizes and by what it identifies as manipulable. Its educational translation is correspondingly modest and probabilistic: if encounter can contribute to direction formation, then education can design the conditions under which such encounters become more likely — it can guarantee exposure, never crystallization. The model should therefore be read as a design hypothesis for educational technology, not as a causal account of motivation formation. What such design looks like in practice, and what it deliberately leaves alone, is the work of Section 5.

4. What the decomposition isolates

4.1 A different kind of remainder

Direction resisted distribution, but it remained a component — something a learner comes to have. The decomposition's second function is different in kind. Run to completion, the sort yields no further component; what it leaves exposed is a relation that was never on the component list: answerability — the standing to have to answer for a judgment, introduced in prior work as the non-transferable relation in which the warrant of a judgment remains anchored in a particular agent (Tomita, 2026b). The grammatical contrast carries the conceptual one: one has a direction; one stands in answerability (Figure 1). A component is a property of the learner that pedagogy can act on. Answerability is not a property of the learner at all. It is a relation between the learner and those entitled to ask — which is why it could not have appeared as a fourth box however finely the sorting continued, and why treating it as one more trait to be fostered would repeat, at the level of theory, the conflation this paper exists to undo. The term is used here in a specific sense. Answerability already circulates in education through a Bakhtinian lineage — the answerability of an utterance or act, an ethics of being implicated (e.g., Ewald, 1993); the present sense is Darwall's, the standing to have to answer to one who can demand, and the two should not be elided.

Figure 1 (image omitted). Direction, drive (available activation and motivational drive), and mode are properties the learner has, shown as boxes within the dashed component region. Answerability is not a fourth component but a relation of address and response in which the learner stands: the standing to answer, to those entitled to ask, for a judgment. The figure marks the categorical contrast between components one has and a relation one stands in.

4.2 Not an artifact of the axis

An objection should be met before anything is built on the claim: that the delegability axis was constructed to exclude answerability, so its isolation is trivial — true by definition. Two replies, one procedural and one empirical. Procedurally, the sorting rule was applied only to capacities ordinarily counted as agency — orientation, drive, mode of acting; answerability nowhere figures in the rule, and it surfaces only when the sorted capacities, however completely enumerated, fail to explain who can be called to answer for a judgment. Empirically, the exclusion was not smuggled in by our vocabulary: in one of the most detailed multidimensional treatments of machine agency available, the terms responsibility, accountability, answerability, blame, and delegation do not occur (Dung, 2025) — an absence in that account, not a verdict on the entire field. The axis did not push answerability out of this strand of agency theory. What the absence illustrates is an omission in one detailed capacity-based account, not the absence of answerability across the wider field. The shape of the move repeats the method of prior work: there, scoring what could be scored located what could not be legitimately delegated (Tomita, 2026b); here, sorting what can be delegated exposes what was never up for delegation. The axis is not chosen to exclude answerability; it is chosen because generative AI has made handover the practical question for every component of learning agency.

4.3 The second-personal ground

What kind of fact is answerability, if not a capacity? Here the paper borrows, and the borrowing must be marked precisely. From Darwall (2006) we take the structure of second-personal address. Claims and demands are addressed person to person, and valid address presupposes, as its felicity condition, the addressee's second-personal competence: the capacity to recognize the demand's validity, to hold oneself to it, and so to have to answer to the addresser. Being responsible for something, in this account, is conceptually tied to being responsible to someone. In educational settings this relation is not abstract moral theory but the ordinary structure of being asked to justify one's solution, interpretation, or design choice to another person. What is borrowed is this address structure. What is original here is the extension: that the warrant of a judgment, no less than conduct, requires a bearer who can be addressed and challenged, and who must answer. Darwall theorizes demands on action; the claim that evaluative judgment stands in the same second-personal relation is this paper's move — consonant with his text, but not stated in it. Where Darwall has been brought into education before, it has largely been by way of respect and dignity (Giesinger, 2012); applying his second-personal accountability to the warrant of a learner's judgment, as the concept education must protect under generative AI, is a different route.

One feature of his account, however, transfers with particular force. Darwall freely admits representation on the claiming side: trustees may demand on another's behalf, and third parties may hold reactive attitudes for a victim. The answering side never appears by proxy — guilt's natural expressions are confession, apology, and self-addressed reproach. On the reading developed here, that asymmetry marks the joint: what cannot be done on one's behalf is precisely to answer. This sharpens a distinction prior work had begun to draw: where accountability can diffuse across institutional structures, answerability names the non-transferable relation in which the warrant of a judgment stays anchored in a particular agent (Tomita, 2026b). This article carries that distinction into education — reserving answerability for the non-transferable standing just described, and keeping accountability for institutional relations, which can, by design, be delegated, audited, and reassigned. This standing runs in the opposite direction to the authority that Xing et al. (2026) find learners conferring on AI teachable agents: authority is conferred by the learner and tracks perceived competence, whereas standing is retained by the one who judges and tracks answerability. The agent can be granted more authority without acquiring any standing.

Recent work offers a complementary route from the AI side: Archer and Wiltsche (2026) argue from systems' cognitive limits — the absence of world-directed, temporal, and anticipatory cognition — through an epistemic answerability to the world, that responsibility lies not in the model but in the system surrounding it, and cannot be offloaded. The two routes divide the labor — that account explains the systems, this one locates the relation — and the division is principled: an argument from present limits must be re-made each time capability moves; an argument from the structure of address does not.

4.4 What the relation grounds

Three matters must not be conflated: answerability as a relational standing; answer-giving competence as a developable capacity; and answerable practice as the observable enactment of that standing. Institutions assign or recognize the standing, pedagogy develops the competence to occupy it well, and research can observe its institutional allocation and behavioral enactment.

The second layer is not a decoration on the first; it is what keeps the first from collapsing into capability education. Strong, coherent direction guarantees nothing by itself — a skilled fraud has direction in abundance, and answers to no one until made to. The qualification is the point: a human fraud can be made to answer, and that very susceptibility — to being addressed, challenged, and held — is answerable practice: the observable enactment of a standing one occupies. A current AI system cannot be made to answer in this sense: nothing can be taken from it, no commitment of its own is at stake, and self-reproach, were it scripted, would idle. It has no such standing; for an AI-assisted judgment, the answer is institutionally assigned to a human agent. The difference is not one of degree along any of Section 2's dimensions; it is the difference between refusing a relation and being unable to stand in one. The asymmetry has a practical edge the other way as well: uncritical delegation does not leave the learner's side untouched: shifts in the frequency and sequencing of planning, monitoring, and evaluation under heavy AI reliance (Fan et al., 2025) motivate concern that the practices sustaining answerable judgment may weaken, though they do not demonstrate erosion of answerability itself — which is why the prescription that follows treats delegation as something to be shaped, not merely permitted.

The boundary this draws does not turn on a metaphysical verdict about machines. It does not require settling whether future systems may simulate address-like behavior; it concerns the educational institution's present criterion for standing — whether a learner can be called to own, revise, and answer for a judgment as theirs. That learners may experience AI systems as second-personal interlocutors — consulting them, trusting them, calling them partners — is consistent with well-attested general mechanisms of anthropomorphism and mind perception (Epley et al., 2007; Gray et al., 2007); its prevalence in educational settings remains an empirical question, and the experience does not imply that the system stands in the relation; perceiving the system as a partner and its being a legitimate addressee of answerability are distinct, on the same reading that separated conferred authority from retained standing above.

On this extension, Darwall's structural point secures the gap: the second-personal notions — authority, valid claim, second-personal reason, accountability — form an interdefinable circle that cannot be entered from outside it, so no enumeration of first-layer capacities, however long, adds up to entry.

This, in turn, grounds what prior work could only state as a condition. (By stipulation, this article reserves accountability for institutionally allocatable duties — reportable, transferable, auditable — and answerability for the non-transferable standing; Darwall's own use of the former term is broader.) Tomita (2026b) held that the delegation boundary stands as long as legitimacy over value judgments remains institutionally attributed to humans — a formulation that leaves the attribution looking like convention. It is not. Institutions keep attributing judgment to humans because the standing is relationally and institutionally assigned, and second-personal competence is required to occupy it validly; the condition is backed by structure, and would have to be re-engineered, not merely re-decided, to move. Nor does the argument lean on the inner opacity of present systems, so it is robust to progress in interpretability: however legible a system's internals become, legibility yields evidence about capability, never standing. And nothing here forecloses the question whether some future system could acquire second-personal competence — any movement of the boundary would require both credible second-personal competence and a legitimate assignment of standing; what the argument forecloses is shortcutting that assignment by capability alone.

5. The two-layer prescription

The two-layer structure yields a two-layer prescription, and the layers — though pedagogically concurrent — must not be conflated. Layer one addresses everything the decomposition can place — it maps components, designs encounters, and configures AI so that delegation removes barriers without substituting for the learner. Layer two addresses what layer one cannot reach. For each prescription we state a principle and one worked example, and we say explicitly what is left to others (these residues are collected as a research agenda in Section 6); the restraint is not modesty for its own sake but the framework's own discipline — a paper about non-delegable judgment should not deliver judgments that belong to institutions.

5.1 Layer one: an analytic compass

Principle. The OECD's Learning Compass 2030 placed agency at the center of a framework for how learners navigate an uncertain world, but it is a compass that points a direction to travel, not one that resolves that direction into components an educator can read (OECD, 2019). This layer supplies a complement: an analytic compass that disaggregates the direction the OECD's compass indicates. Read a learner's components separately instead of guessing at aptitude as a single quantity. Direction, available activation, motivational drive, and mode are distinct analytic readings, and several have established instruments adjacent to them — self-regulated-learning measures speak to mode and its trainability (Zimmerman, 2000), engagement and vitality measures to available activation, interest inventories to candidate directions. The compass is a reading frame for existing instruments, organized by the delegability sort rather than summed; the reliability and validity of this mapping require separate construct-validation research. One reservation is constitutive rather than cautionary: the compass reads only what can be externalized. Answerability should not be represented as an individual trait score, and a compass that claimed to score it would re-import the conflation this paper identifies. Its institutional allocation and observable enactment can nevertheless be studied — who is entitled to ask, who is required to respond, who actually gives reasons; what remains inaccessible to direct measurement is sincere endorsement. This is the point at which both the OECD's compass of direction and this compass of components reach their limit. This is also the precise sense in which the present account accepts the demand that agency remain measurable rather than a residual aspiration (Cukurova, 2026) — and then locates, at that demand's edge, the standing itself; the competence to occupy it well is developed, second-personally, in layer two.

Worked example. A learner presents as "unmotivated." The compass disaggregates: available activation high — the same student is animated elsewhere; direction absent; mode predominantly critical. The reading is not a motivation deficit but a pre-crystallization state with a strong engine and no object, and the indicated lever is exposure to encounters, not exhortation. One further effect is rhetorical but consequential: the compass replaces find your purpose — a binary demand that mostly paralyzes — with observe your own gradients, an incremental practice a pre-crystallization learner can actually perform.

Left to others. Which instruments, with what weights, against which values: to each institution, whose reading of its own learners is itself a judgment someone must answer for.

5.2 Layer one: encounter engineering

Principle. If encounter can contribute to direction formation (Section 3), then curricula can be designed for encounter density: structured contact with problem-holders, early and embodied, before competence would conventionally "justify" it. Medicine has long operated a cognate design: clinical clerkship routes students through patients on the premise that direction toward care is formed at the bedside rather than installed in the lecture hall — a premise institutionalized by practice, not causally verified in this article's terms. The model can serve as a design heuristic wherever curricula can create sustained contact with problem-holders.

Worked example. For disciplines without wards, the encounter can be miniaturized into what we call a discomfort inventory exercise: the learner delegates a task wholly to an AI system, then inventories every discomfort with the output. The inventory is the learner's first specification profile — the felt difference between fluent output and right output is exactly where tacit commitment surfaces, following the ablation logic documented in Tomita (2026a). The exercise turns the AI's weakness-without-specification into a mirror, and a sequence of such micro-encounters into a curriculum. Independent evidence is consistent with the design: studying EFL learners revising AI-drafted statements of purpose, Jiang et al. (2026) found revision patterns ranging from compliance-oriented acceptance through form-oriented modification to content-oriented innovation — a gradient this article reinterprets as degrees of crystallized direction, a reading the authors themselves do not make — with two factors dividing them: the felt gap between an authentic and a GenAI-constructed voice, and enthusiasm to present one's intentions. Those two map onto components this article decomposed independently — what this article tentatively reads as a precursor of answerable practice (a judgment one will have to own), and the motivational drive of a direction one has taken up — surfacing as empirical dividing factors in a study that did not set out to test the decomposition. Where Jiang et al. document the gradient and prescribe that learners "remain the agents," the present account provides a design conjecture for what their study observed after the fact: the discomfort inventory is designed to elicit conditions hypothesized to support movement along that gradient; the proposition remains untested.

An ethical boundary. Encounter engineering must not instrumentalize the people encountered. Problem-holders are not teaching materials: consent, reciprocity of benefit, protection from repeated emotional labor, and structured reflection are design requirements, not optional refinements.

Left to others. Which problem-holders, at which sites, with what safeguards: discipline-specific terrain, owned by those who know it. The design guarantees exposure, never crystallization, and an honest program says so.

5.3 Layer one: AI guardrails

A prior distinction. Performance delegation and developmental delegation are different objects: offloading production can raise artifact quality while leaving knowledge and transfer unchanged (Fan et al., 2025). What guardrails protect, therefore, is not the artifact but the episodes in which the learner answers for judgments. Principle. Configure educational AI to distinguish two requests that look alike. When a prompt carries the learner's specification — their constraints, judgments, reasons — completion may primarily reduce externalization cost; whether it is educationally legitimate still depends on the learning objective, the learner's developmental stage, the plan for fading support, and the criterion for unaided transfer. When the prompt is specification-empty, completion would substitute for the very thing education exists to develop; the right behavior is to route the request back into dialogue until the learner's own commitments appear, and only then to draft.

Worked example. An essay assistant receives "write my essay on X." Instead of producing text, it asks what the learner finds wrong with the standard view of X, which two sources they would trust and why, and what conclusion they would defend if pressed — and drafts only from the answers. The same model, the same capability, a different default.

Left to others — with two stated limits. Detection of specification can be gamed; a determined learner can launder emptiness through borrowed opinions, so the guardrail shapes the default path rather than guaranteeing the outcome. And dialogue with an AI, however well configured, does not cultivate answerability — the system can elicit commitments, but it cannot be the one to whom the learner answers. Guardrails serve layer one. They do not substitute for layer two.

5.4 Layer two: the second person

Principle. The standing of answerability is conferred by the relation itself; what grows is the learner's competence to answer well, and that competence, however rehearsed, is normatively tested and sustained only in second-personal exchange: being questioned, answering, and owning the answer. The educator's irreplaceable function — the one no configuration of tools absorbs — is to be the other person: to ask the learner what is your assessment? and to hold them to what they say. AI can help learners rehearse this competence — drafting reasons, anticipating objections, simulating questions — and can mediate the exchange; what it cannot occupy is the position of final addressee, the person to whom the answer is owed. And the entitlement to ask is not innocent of power: a relation in which only the learner answers degenerates into interrogation. The educator who asks must also be answerable — for the questions posed, the criteria applied, and the decisions taken on the learner's work. Nor does answering privilege one performance format: reasons can be given in writing, multimodally, or with preparation, so that the standing to answer is not confounded with rapid oral fluency in a particular language or neurotype. Clinical supervision often exemplifies this structure: once responsibility for a patient-facing judgment is institutionally assigned, the trainee stands answerable for it; what training builds is the competence to answer well, and it builds it by making the trainee answer, case after case, to someone entitled to ask. The structure generalizes beyond medicine: a learner's standing does not wait for graduation — from the moment work is submitted under their own name, the answer is theirs — and what education builds is not the standing but the competence to occupy it. Assessment forms that stage this exchange — oral defense of one's own work, process accounts the learner must stand behind — are implications worth one sentence each here and a literature of their own elsewhere.

The prescription is structurally grounded, not nostalgic. When AI accelerates the externalization and combination of knowledge, the rate-limiting processes shift to socialization and internalization (Nonaka & Takeuchi, 1995; Tomita, 2026b) — and socialization, where evaluative commitments are contested and revised between persons, is precisely the second-personal site. The relational conditions that self-determination theory found necessary for internalization point the same way (Ryan & Deci, 2000). Nor is this a retreat from human–AI configurations: the hybrid is the right vessel, and layer one fills it. But within any configuration, the answerable party is the human in it, not the configuration — layers that mesh, not a fusion that answers. What layer two restores is older than the lecture hall: apprenticeship's core, with its drudgery finally delegable, returned to the center.

Left to others. Teacher development and assessment regimes that institutionalize the exchange: stated here as implications, designed by those who will answer for them.

6. Boundary conditions and research agenda

What this article leaves to others is collected here as testable commitments rather than scattered disclaimers. The encounter–crystallization–conversion model is a design conjecture, and the two-layer structure yields propositions that could falsify or refine it. P1. AI drafting support may improve assisted artifact quality without comparable gains in unaided knowledge or transfer (cf. Fan et al., 2025); whether adding structured answerability episodes improves subsequent transfer is a separate moderation hypothesis to be tested. P2. Encounter-rich curricula will produce stronger observable indicators compatible with endorsed direction — voluntary re-engagement, persistence across changes in supervision, reasoned defense across contexts — than information-rich curricula of equivalent coverage; no single indicator is treated as proof of sincere endorsement. P3. Enacted answerability episodes — rather than formal allocation alone — will predict deeper revision, error correction, and justification, after controlling for time-on-task and prior attainment. P4. The allocation and enactment of answerability are observable through questioning episodes, reason-giving, revision records, and decision ownership; divergence between recorded and enacted answerability is an analytic indicator, and sincere endorsement is not directly inferable from compliance. The decomposition also reads excess without treating activity level as a diagnosis. In occupational research, work engagement and workaholism both present as high activity while differing in motivational quality (Schaufeli et al., 2006); used here as a boundary-case analogy, that contrast prompts two inquiries — whether available activation is being depleted by ill health or adverse circumstance even while the direction remains endorsed, and whether the direction anchoring the activity is endorsed at all or compulsively held. "Just stop" is no more a component-level intervention than "just start": instruction alone neither restores available activation nor resolves what anchors the direction; and because sincere endorsement cannot be inferred from activity level, slowing a learner down remains a relational act — performed by someone entitled to ask, who in turn owes the learner reasons for the stop. Boundary conditions equally belong here: the framework is developed for formal education with identifiable educators. Its extension to informal, massive, or fully automated settings, the accessibility of reason-giving formats across languages and neurotypes, and the ethics of encounter designs involving real problem-holders are left open — as a research agenda rather than a residue.

7. Conclusion

What remains for the learner when AI can do the work? The question dissolves once agency is decomposed along the axis the question itself supplies. Many functions gathered under agency may be supported or partially offloaded when the design preserves development, transfer, error detection, and independent judgment — the sort in Section 2 says where, and layer one of the prescription says how. What cannot be handed over is of two kinds, and keeping them distinct is this article's central discipline: direction, a component whose endorsement cannot be substituted and may be supported by encounter; and answerability, not a component at all but the standing in which a learner answers to someone for a judgment. AI changes the economics of everything around that relation, stripping the externalization burden from teaching and learning alike; what it does not change is the relation itself: who must answer, and to whom. The educator's position is thereby clarified, not diminished: design the encounters, configure the tools, read the compass; and then ensure that the learner remains answerable to a legitimate human addressee — and, often, be that person. Education in the age of AI is not the defense of human tasks against automation. It is the cultivation of what was never a task.

Practitioner Notes

What is already known about this topic - Generative AI systems now plan, draft, and revise, making learner agency a central contested concept in educational technology. - Distribution accounts treat agency as a property of human–AI systems; psychological accounts treat it as an individual capacity to be fostered. The two run in parallel.

What this paper adds - A decomposition of learner agency by delegability — direction, two forms of drive, and mode — rather than by psychological function. - A conjectured sequence of direction formation: embodied encounter may support the crystallization of commitment and the anchoring of available activation as motivational drive. - Answerability identified as a relation isolated, not contained, by the decomposition: the non-delegable standing to have to answer for a judgment. - A two-layer educational prescription matching the two layers.

Implications for practice and/or policy - Read a learner's components separately, as an analytic map rather than a high-stakes diagnostic classification, instead of guessing at aptitude as a single quantity. - Include ethically designed encounters with consequential problems and affected persons rather than relying on exhortation alone. - Configure educational AI to route specification-empty requests back into dialogue rather than completing them. - Protect second-personal exchange — being questioned, answering, owning the answer — as the site where answerability is enacted and answer-giving competence develops. (Answerability here means that learners must remain able to say why a judgment is theirs when questioned by another person.)

Statements

Conflict of Interest. The author declares no competing interests.

Data Availability. This is a theoretical article. No datasets were generated or analyzed. All sources discussed are publicly available and cited in the references.

Generative AI Use. The author used generative AI tools (large language models) as drafting and editing aids during manuscript preparation, and as an object of analysis discussed in the text. All conceptual claims, the framework, the argument, and final wording are the author's own; the author takes full responsibility for the content.

Funding. None

References

Anthropic. (2026, April 20). Meet the winners of our Built with Opus 4.6 Claude Code hackathon [Blog post]. https://claude.com/blog/meet-the-winners-of-our-built-with-opus-4-6-claude-code-hackathon

Archer, K., & Wiltsche, H. (2026). The origins of artificial intelligence in natural intelligence [Preprint]. PhilArchive. https://philarchive.org/rec/ARCTOO-2

Bandura, A. (2001). Social cognitive theory: An agentic perspective. Annual Review of Psychology, 52, 1–26. https://doi.org/10.1146/annurev.psych.52.1.1

Cathcart, S., Priestley, M., Priestley, A., & Rushton, E. A. C. (2026). Unravelling agency: A critical review of conceptualisations of agency for educational research, policy and practice. Review of Education, 14(1), e70131. https://doi.org/10.1002/rev3.70131

Cukurova, M. (2026). Agency as a system property in human–AI interaction in education. British Journal of Educational Technology, 57(4), 1065–1070. https://doi.org/10.1111/bjet.70060

Darwall, S. (2006). The second-person standpoint: Morality, respect, and accountability. Harvard University Press.

Dung, L. (2025). Understanding artificial agency. The Philosophical Quarterly, 75(2), 450–472. https://doi.org/10.1093/pq/pqae010

Emirbayer, M., & Mische, A. (1998). What is agency? American Journal of Sociology, 103(4), 962–1023. https://doi.org/10.1086/231294

Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864

Ewald, H. R. (1993). Waiting for answerability: Bakhtin and composition studies. College Composition and Communication, 44(3), 331–348. https://doi.org/10.2307/358987

Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., & Gašević, D. (2025). Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology, 56(2), 489–530. https://doi.org/10.1111/bjet.13544

Giesinger, J. (2012). Respect in education. Journal of Philosophy of Education, 46(1), 100–112. https://doi.org/10.1111/j.1467-9752.2011.00836.x

Gray, H. M., Gray, K., & Wegner, D. M. (2007). Dimensions of mind perception. Science, 315(5812), 619. https://doi.org/10.1126/science.1134475

Hidi, S., & Renninger, K. A. (2006). The four-phase model of interest development. Educational Psychologist, 41(2), 111–127. https://doi.org/10.1207/s15326985ep4102_4

Jiang, Y., Wu, Q., Yang, Y., Jian, C., & Zhao, J. (2026). Learner agency in revising GenAI-generated statements of purpose. British Journal of Educational Technology, 57(4), 965–983. https://doi.org/10.1111/bjet.70041

Järvelä, S., & Hadwin, A. F. (2013). New frontiers: Regulating learning in CSCL. Educational Psychologist, 48(1), 25–39. https://doi.org/10.1080/00461520.2012.748006

Kortemeyer, G., Dittmann-Domenichini, N., Schlienger, C., Spilling, E., Yaroshchuk, A., & Dissertori, G. (2023). Attending lectures in person, hybrid or online—How do students choose, and what about the outcome? International Journal of Educational Technology in Higher Education, 20(1), 19. https://doi.org/10.1186/s41239-023-00387-5

Luo, J. (2025). How does GenAI affect trust in teacher–student relationships? Insights from students' assessment experiences. Teaching in Higher Education, 30(4), 991–1006. https://doi.org/10.1080/13562517.2024.2341005

Nonaka, I., & Takeuchi, H. (1995). The knowledge-creating company: How Japanese companies create the dynamics of innovation. Oxford University Press.

OECD. (2019). OECD learning compass 2030: A conceptual learning framework [Concept note]. OECD. https://www.oecd.org/content/dam/oecd/en/about/projects/edu/education-2040/concept-notes/OECD_Learning_Compass_2030_concept_note.pdf

Ryan, R. M., & Deci, E. L. (2000). Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist, 55(1), 68–78. https://doi.org/10.1037/0003-066X.55.1.68

Scardamalia, M. (2002). Collective cognitive responsibility for the advancement of knowledge. In B. Smith (Ed.), Liberal education in a knowledge society (pp. 67–98). Open Court.

Schaufeli, W. B., Taris, T. W., & Bakker, A. B. (2006). Dr Jekyll or Mr Hyde? On the differences between work engagement and workaholism. In R. J. Burke (Ed.), Research companion to working time and work addiction (pp. 193–217). Edward Elgar.

Tomita, K. (2026a). Domain-native development: A Mekiki framework for AI-assisted knowledge work [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/cwkav_v1

Tomita, K. (2026b). Philosophy as cognitive assay: Measuring the delegation legitimacy boundary in AI-assisted knowledge work [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/e9qw5_v2

Xing, W., Kim, T., Song, Y., Li, H., Li, C., & Kim, J. (2026). Unveiling interaction patterns between students and a generative AI teachable agent: Focusing on students' agency and AI agents' authority. British Journal of Educational Technology, 57(4), 896–923. https://doi.org/10.1111/bjet.70038

Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2

Zimmerman, B. J. (2000). Attaining self-regulation: A social cognitive perspective. In M. Boekaerts, P. R. Pintrich, & M. Zeidner (Eds.), Handbook of self-regulation (pp. 13–39). Academic Press. https://doi.org/10.1016/B978-012109890-2/50031-7