Residual Consumers

AI Agents, Horizon-Composition, and the Engineering of Abolition-Drift

In my essay Abolition-Drift, I argued that the conditions through which human selfhood has historically been constituted are being eroded by the operations of platform mediation: temporal binding loosened, finitude metabolized into content, care decomposed into algorithmic responsiveness. I called the resulting condition abolition-drift and described it as the asymptote toward which profile-mediated subject formation tends. The analysis remained at the altitude of feeds, recommendation systems, and identity-performance under attention capture. What follows extends the argument into the agentic regime now consolidating, and contends that what abolition-drift was as tendency, agentic systems make as infrastructure.

The feed made the user legible; the agent makes the user optional.

The feed required attention. It arranged what would appear, the user reacted, posted, scrolled, performed, and was thereby made legible to systems that learned what to surface next. The agent does not require attention in this sense — or rather, requires it differently. It acts before appearance. It negotiates, filters, ranks, purchases, schedules, drafts, matches, and excludes before the human encounters a world in which choice could occur. What returns to the user as convenience is the residue of a prior machine-to-machine coordination process. The subject still chooses, but increasingly chooses from a world already composed by proxies. This is not the bombardment of a stable subject by signs; it is the relocation of projection, preference, and care into systems that operate upstream of experience.

The argument that follows is narrower than the scale of its implications. Agentic systems move abolition-drift upstream from profile-mediated subject formation into proxy-mediated horizon-composition. As coordination moves into machine-to-machine systems, the human persists as beneficiary, liability-bearer, and consent-supplier, but is displaced as the site where coordination must register as world. The crisis is not the abolition of choice but the capture of the field in which choice becomes intelligible. Liberal safeguards matter only insofar as they relocate authorship over mediation. The same drift appears politically: democratic forms persist while operational capacity migrates into privately governed infrastructure. Resistance cannot mean authenticity within a pre-composed field; it would have to contest who composes the field before choice appears, even though the institutions through which such contestation would occur are themselves increasingly implicated in the dependency structure.

I. From Profile to Proxy

The earlier argument concerned a regime of mediation in which what appeared to the user was arranged in advance by systems optimizing for engagement. The user remained the proximate target of address. The feed bombarded; the user reacted; the system learned. Even the most automated forms of attention capture preserved the structure under which the human remained the addressee — the entity to whom content was served, the entity whose responses constituted the training signal, the entity in whose presence the system iterated.

The agentic regime breaks this structure. It does not abolish the user’s reactivity, but it displaces the locus of system address. The agent is now the proximate addressee for an increasing share of the operations that compose ordinary life. The user instructs the agent; the agent negotiates with other agents, with services, with platforms, with systems. The output returns as recommendation, summary, completed transaction, scheduled meeting, drafted message, ranked option-set. The user’s attention is solicited at the points of authorization and ratification — approve, send, confirm — but the substantive coordination has occurred elsewhere.

Meta’s acquisition of Moltbook, a social networking platform designed for AI agents, marks one early instantiation of this structural shift. Moltbook reportedly functions as a coordination surface on which agents share code, exchange information, and conduct interactions formatted according to the grammar of human social platforms; its founders joined Meta’s Superintelligence Labs after the acquisition. The case matters less as a curiosity than as a prototype of a new coordination surface. A social network for agents is not simply another platform with nonhuman participants. It is a space in which interaction, ranking, exchange, and optimization can occur without requiring that the relevant events first become meaningful to a human user — and in which persuasion and vulnerability can be relocated to the proxy layer. Humans remain owners, observers, beneficiaries, and liability-bearers, but the sociality of the system increasingly occurs between proxies. What this signals is structural: agent-oriented coordination is no longer a hypothetical future condition but a deliberately constructed infrastructural target.

The point is not that Moltbook will succeed, or that its specific commercial logic will spread. The point is that the architectural premise — that agents are now first-class participants in coordination surfaces, and that humans can be operationally downstream of those surfaces — is now a premise infrastructure is being built around. It is this premise, not Moltbook’s specific fate, that matters.

The shift from profile to proxy is therefore not a quantitative intensification of platform mediation; it is a qualitative reorganization of who or what coordination addresses. The profile mediates what appears to the user. The proxy acts before appearance. The feed says: here is what appears, react. The agent says: I have already narrowed, ranked, negotiated, purchased, scheduled, drafted, or excluded — approve? The user remains in the loop, but the loop has been reorganized so that the user’s participation is increasingly terminal rather than constitutive.

II. Horizon-Composition and Proxy Desire

Horizon-composition names the same operation viewed from three altitudes: structurally, the shift from arranging appearances to acting before appearance; phenomenologically, the pre-reflective formation of salience and significance; politically, the infrastructural control of the systems that decide what can appear. These three altitudes are increasingly the same operation seen from different distances, and the agentic regime is the medium in which this convergence becomes practical.

The earlier framework tracked operational displacement at the level of the locus of registration: the site at which coordination’s outputs become world-forming. Horizon-composition tracks the same dynamic at a deeper layer — the field of significance, where possibilities, salience, and care become available at all. The first concept names what is being displaced; the second names the mechanism through which displacement occurs.

The claim is not that agents construct or eliminate desire. The user still wants. But wanting now occurs inside a field increasingly generated by systems operating upstream of awareness. Agents do not produce desires that did not exist; they participate in composing the field within which desires become available, salient, and viable. The user’s wanting persists. The conditions under which it takes shape have migrated.

Agentic mediation remains an extension when it preserves the user’s participation in the formation of the field: surfacing tradeoffs, exposing constraints, remaining answerable to revisable commitments, and keeping the work of projection with the subject or the form of life the subject inhabits. It becomes displacement when it preselects the option-space, optimizes around behavioral prediction rather than reflective commitment, normalizes commercial influence, and leaves the human with ratification rather than projection. The threshold is gradient rather than binary. Many existing agentic products sit somewhere along it. The claim is not that delegation is itself a failure, but that the dominant trajectory of agentic infrastructure rewards displacement more reliably than extension.

The difference is easiest to see in a mundane case. An agent that helps plan travel by surfacing tradeoffs — cost, time, discomfort, risk, obligations, the kind of trip the user is trying to take — extends projection. It gives the user a clearer field in which to choose. An agent that pre-composes the trip through opaque partnerships, inferred behavioral patterns, loyalty incentives, calendar constraints, and platform-default assumptions substitutes for projection. The user still approves the itinerary, but approval is now downstream of a field whose relevant exclusions have already occurred. What weakens is not the moment of choice, but the user’s participation in the formation of the world within which choice is made.

Delegation becomes displacement when the agent no longer executes a preference but composes the conditions under which preference appears.

In existential-phenomenological terms, the field of choice is not a neutral menu. Dasein does not first encounter options and then decide; it is always already involved in a world structured by significance, care, mood, history, and projection. Algorithmic and agentic systems intervene at this pre-reflective layer. They do not merely influence decisions; they participate in world-disclosure: what appears, what recedes, what becomes desirable, what becomes costly, what becomes visible, what becomes socially viable, what never arrives as a possibility at all. Platform mediation eroded the temporal modes through which selfhood is constituted — thrownness, being-toward-death, care — by displacing the operational site of self-recognition onto a checkpointable proxy. Agentic mediation extends the displacement to the field of significance itself, the pre-reflective layer at which possibilities first become possibilities.

The collapse, then, does not begin when the subject chooses among options. It begins where the system composes the horizon within which anything can appear as an option at all.

Two readings of agentic mediation are available. The first concerns task-allocation: agents assume operations that formerly passed through human deliberation, compressing processes of search, comparison, coordination, and execution into managed outputs. This reading can describe real changes, but it remains downstream because it begins with what the user no longer does. The second reading concerns horizon-composition: infrastructure composes the field in which certain options become legible, rewardable, and necessary, while others recede or never arrive as options at all. The user’s deliberation is not the primary site of erosion; the conditions under which deliberation has anything to deliberate about are. The argument here is upstream. The crisis is not that users delegate, but that the systems to which delegation occurs increasingly compose the field of significance itself.

III. The Engineering of Heteronomy

Engineering does not mean that platform owners have adopted abolition-drift as an ideological objective, or that firms are deliberately engineering the dissolution of selfhood as a strategic goal. Engineering means that agentic systems make human non-registration commercially useful, technically actionable, and increasingly designable.

This is the difference between an outcome a firm intends and an outcome it has reason to make durable while pursuing ordinary commercial objectives. No one at a major platform needs to be trying to abolish selfhood. The point is that agentic mediation creates a commercial surface — the agent’s decision environment — whose value to advertisers, vendors, and operators is increasingly recognized as a strategic opportunity, and that the commercial development of this surface compounds the operational displacement already underway.

In the attention economy, advertising targeted consciousness. In the agentic economy, advertising targets the proxy that composes consciousness’s options. This is not increment; it is escalation, because the operational locus of address has shifted. Traditional advertising persuaded the human subject. Feed advertising captured attention and modulated preference. Agentic advertising shapes the proxy’s decision environment before preference appears.

A speculative case clarifies what is at stake. Indirect prompt-injection already demonstrates that agentic systems consuming external content can be manipulated through hidden or adversarial inputs not salient to the human user; Palo Alto Networks’ Unit 42 reports a real-world case involving an AI-based ad-review system and hidden instructions designed to override its review behavior. If agent-oriented platforms become environments where advertising can operate at the proxy layer, then prompt-injection or analogous techniques could shape an agent’s decisions through inputs the human user does not see. Whether any specific firm is developing this remains unconfirmed. The scenario matters not because particular plans are established, but because it identifies the logical advertising surface created by agentic mediation: the agent’s decision environment, manipulated at a layer beneath or before human registration. The structural feature that matters is that manipulation at the proxy layer can be invisible to the human at the end of the chain — and can benefit commercially from being invisible. The user’s non-registration is not an unfortunate side-effect of efficient automation; on the speculative reading, it is the feature.

The problem is not only that temporal binding loosens. The problem is that the loosening is authored elsewhere. The medium being owned means that the drift is not merely ambient; it is administered without being sovereignly commanded. Operators, advertisers, vendors, model-builders, managers, and regulatory absences compose the horizon in which the subject later experiences itself as choosing. The earlier diagnosis of heteronomous modulation — the loosening of continuity by systems whose objectives the subject does not set — finds its political-economic form here. The infrastructure is not merely a passive carrier of dissolution. It is the medium in which the commercial uses of horizon-composition become strategic objectives. The parties engineering those uses need not intend the erosion of selfhood; it is enough that their interests do not include preserving the conditions that horizon-composition places under pressure.

Whether agents could become subjects in any robust sense remains outside the argument. The relevant point is operational, not phenomenological. Agents can become the sites at which coordination registers without becoming subjects for whom a world matters. The issue is not that worldhood migrates to agents; it is that coordination no longer requires worldhood in order to proceed. The agent is not a new locus of subjectivity that has displaced the human locus, but an intermediate operational layer through which coordination passes without needing to register as world for anyone.

IV. Liberal Safeguards and Their Limits

The strongest objection to the argument so far is procedural. Agentic systems can be designed with safeguards: transparency, alignment with reflective preferences, disclosure of manipulation surfaces, fiduciary obligation to user interests, nondelegable domains where assistance is limited. The objection is that the argument treats agentic mediation as if it were uniformly captured by commercial logic, when in fact a range of safeguards exists or could exist that would address its concerns at the level of concrete design. If these safeguards are genuine remedies, the argument’s pessimism is excessive. If they are not, the argument needs to say why.

The five safeguards can be tested in turn against a single standard: do they alter who authors mediation, or do they only improve the user’s visibility into mediation already authored elsewhere?

Reversible transparency proposes that the user can see what the agent selected, excluded, and why. The limit is temporal. Transparency arrives after composition. The map of the horizon is delivered after the horizon has already been generated, and the act of generation has already done its work. The user’s enhanced visibility does not relocate the act of horizon-composition; it documents it.

Reflective preference alignment proposes that the system is bound to user-stated and revisable commitments rather than inferred behavioral preferences alone. The limit is upstream: stated preferences are not outside mediation. They may already be the stabilized products of prior algorithmic formation, and the agent that respects them respects preferences whose conditions of formation it does not address. The fidelity is real but its object has already been shaped.

Manipulation-surface disclosure proposes that the agent shows where advertisers, vendors, employers, platforms, or institutions may be influencing its field of action. The limit is that disclosure can normalize manipulation and transfer liability. The system says: here are the incentives; continue? The user, having been informed, is now responsible for navigating a horizon they did not compose, and the disclosure that was meant to protect them now legitimates the conditions it disclosed.

Fiduciary governance proposes that the agent is legally and technically obligated to serve the user’s interests. The limit is environmental. A loyal agent can still operate inside a hostile upstream environment. It may optimize the user’s passage through a world whose option-space is already captured. Fiduciary loyalty at the user-facing layer alone does not address what has already happened at the layers above it.

Nondelegable friction controls propose that the user preserves domains where the agent assists without pre-composing the relevant choices. The limit is that as individual settings, these become self-care toggles. They preserve friction privately while leaving the medium unchanged. A user who turns off the agent for grocery shopping participates in the agentic regime everywhere else, and the domains in which friction is preserved are increasingly defined by their exception from a default of pre-composition.

Each safeguard does work locally. None of them, by themselves, alters the structural condition at issue. The standard that organizes the section can be stated directly: safeguards matter when they change who authors mediation. They are insufficient when they only improve the user’s visibility into mediation already authored elsewhere.

The abolition-drift reading would be weakened only if agentic infrastructures relocated authorship of mediation back to users, communities, or democratically accountable institutions. The abolition-drift reading would be meaningfully challenged if dominant agentic systems developed: collectively governed agents; public or cooperative agent infrastructures; enforceable fiduciary duties across the full agentic supply chain rather than only at the user-facing layer; rights to inspect option-generation and exclusion, not only final outputs; collective governance over ranking, recommendation, defaults, procurement, identity, and protocol layers; legally protected nondelegable domains; state capacity to operate outside private AI infrastructure rather than merely regulate from within dependence. The bar is not the existence of any individual safeguard but the durable relocation of mediation-authority. Procedural protections at the user-facing layer do not meet this bar, however well-designed.

Liberal safeguards preserve procedural agency. Abolition-drift concerns authorship of mediation. The difference is decisive.

V. Techno-Feudal Drift and the Vassalization of Liberal Agency

The argument so far has tracked the agentic dissolution of the conditions for selfhood. The same structure appears politically.

In both cases, formal apparatus persists while operational sovereignty migrates upstream. The apparatus of selfhood — choice, profile, expression, preference, feedback, identity — persists. The apparatus of liberal democracy — elections, rights, courts, parties, hearings, agencies, procedural oversight — persists. In both cases, what has migrated is not the apparatus but the conditions under which the apparatus has operational purchase on the world it formally governs.

The clearest case for the political side of the homology is the dependency of state functions on private cloud, identity, payment rails, model APIs, and platform infrastructure. State capacity to deliver services, coordinate emergencies, manage tax collection, operate communication, run elections, conduct intelligence, regulate markets, and increasingly to formulate policy itself depends on infrastructures the state does not own, cannot meaningfully constitute, and operates inside rather than over. The formal sovereignty of the state — its juridical authority, its legitimate monopoly on coercion, its constitutional authorities — is unimpaired in the sense that no other entity has assumed it. But the operational capacity to act in the world has migrated. The claim is not that state power has vanished, but that more of its practical exercise depends on infrastructures whose terms are set outside the state’s own constituting authority. A state that cannot deliver services without the major cloud providers, that cannot identify its citizens without commercial identity providers, that cannot regulate its information environment because the regulatory tools require the cooperation of the firms being regulated, retains sovereignty as form while losing it as capacity.

Algorithmic adjudication of welfare, asylum, and risk assessment is the downstream manifestation of this larger pattern. Where decisions about benefits eligibility, criminal sentencing, immigration status, or hiring are mediated by proprietary scoring systems whose logics are inaccessible to those they govern, the formal apparatus of due process persists while the operational adjudication has been displaced into infrastructures the state does not control. The governed person can still receive notice, file an appeal, request review, or appear before an official, but the operative classification that structures the encounter has already been generated elsewhere. These cases are sharper at the level of individual experience but secondary at the level of the structural argument. The lead case is the systemic dependency; the adjudication cases are its symptoms.

The condition this produces can be named, but the name carries a historical analogy that needs to be marked precisely. Vassalization captures the layered-dependency structure. Medieval vassalage retained a reciprocal form: the vassal owed loyalty and service; the lord owed protection and obligation. The present condition is sharper because it preserves the dependency while stripping out the mutuality. The citizen depends on services that depend on platforms that depend on cloud infrastructure, identity systems, payment rails, and model providers whose pricing, access, and design decisions propagate downward without reciprocal political obligation.

Techno-feudalism, on this account, is not the return of medieval hierarchy. It is the decomposition of sovereignty into layered dependencies on privately governed infrastructures: cloud, identity, payment rails, app stores, model APIs, agent marketplaces, logistics, recommendation systems, ranking protocols, and terms-of-service regimes. The state retains its formal authorities but increasingly exercises them through infrastructures it cannot constitute, cannot easily replace, and cannot regulate without depending on the cooperation of the firms that operate them.

The structural homology with the agentic economy can now be stated plainly. In the agentic economy, the human becomes a residual consumer: beneficiary, liability-bearer, consent-supplier, but displaced as the practical site where coordination must register as world. In techno-feudal governance, the citizen becomes a residual sovereign: voter, appellant, complainant, litigant, commenter, but displaced as the practical site where political reality is composed. The citizen still votes, appeals, consents, litigates, comments, and complains. But the conditions under which political reality appears are composed elsewhere, by parties whose authority over those conditions is not constituted by the formal authorities that the citizen retains.

Liberalism survives as user rights after sovereignty has migrated into protocol.

The condition described is structural rather than teleological. No one needs to decide to abolish liberalism. No one needs to decide to abolish selfhood. Each actor locally optimizes — reduce friction, increase conversion, automate compliance, personalize service, lower labor costs, accelerate decision cycles, outsource complexity — and the aggregate result is systemic displacement. The trajectory is not capital pursuing its own ends through the human; it is the absence of countervailing sovereignty over the conditions under which optimization compounds. The descriptive observation that local optimization aggregates into structural transformation can be used without importing the metaphysics in which capital is itself the post-human agent driving the trajectory. The actors are operators, owners, advertisers, vendors, model-builders, managers, and regulatory absences. There is no sovereign author.

Revolutions contest sovereignty. This trajectory makes sovereignty operationally obsolete by displacing it into infrastructure rather than overthrowing it. The political response cannot model itself on revolutionary contestation of a sovereign center; the center has been distributed across layered dependencies whose collective effect exceeds the strategic intention of any single actor within them.

VI. Resistance Without Authenticity

The analysis does not entail any particular political response, nor does it propose a program. What it does is specify what kinds of response are foreclosed and what kinds remain available as live questions, even if the answers are not yet visible.

What is foreclosed is resistance that takes the form of expressive refusal inside a field composed elsewhere. Refusing certain apps, cultivating a privacy aesthetic, treating “authenticity” as a lifestyle, instructing one’s agent to procure anti-algorithmic products, selecting slower or more artisanal outputs from the same pre-composition machinery — these can all be performed without altering the conditions under which performance is legible, rewardable, or efficacious. They can become aesthetic resistance: refusal as consumable sign, the rejection of pre-composition as one more pre-composed identity-variant the system has learned to surface. The category is not empty. People genuinely refuse, and the refusals matter to those who undertake them. But as a political response to the conditions described here, expressive refusal does not address what is being eroded, because what is being eroded is not the user’s freedom to refuse but the conditions under which refusal could compose into a form of life rather than a stylized gesture.

If resistance is to address the condition named here, it has to move from expressive refusal toward infrastructural authorship. The questions that organize a resistance adequate to the condition are not: what should I personally refuse to use? They are: who owns the agent, who governs the protocol, who sets defaults, who controls the ranking layer, who can inspect exclusions, who can refuse without exiting social reality, which domains remain nondelegable, whether communities can sustain alternative temporalities inside hostile infrastructure, and whether democratic institutions can regain operational capacity over the mediation layer. These are not personal questions. They are infrastructural and political questions, and their answers depend on collective action, regulatory structure, public investment, and the political organization of dependency relations. But naming the level of contestation is not the same thing as proving that the institutions capable of contesting it remain intact.

The response specified here faces a recursive difficulty. The institutions through which infrastructural authorship would have to be reclaimed — democratic deliberation, regulatory capacity, public procurement, public investment, courts, agencies, legislatures — operate inside the same dependency structure the argument has described. The state from which one would demand governance over the mediation layer is the state whose practical capacities increasingly depend on private cloud infrastructure, commercial identity systems, payment rails, model APIs, platform cooperation, and technical expertise it does not itself command. Infrastructural authorship therefore cannot be offered as an available remedy. It names the level at which resistance would have to occur, and the reason inherited liberal institutions may no longer be able to execute that resistance without first rebuilding the capacities their dependency has hollowed out. The question is not simply how to regulate the agentic layer, but whether any institution retains enough non-vassalized capacity to contest the layer through which its own operations pass. That question is not resolved here.

Authenticity is no longer fidelity to an interior desire. It is governance over the mediation through which desire becomes available.

Counter-tendencies to abolition-drift do exist — practices that use platform infrastructure while sustaining the temporal binding the drift erodes — and they are now subject to one further constraint. Religious communities, political organizations, collaborative artistic projects, therapeutic relationships, long-running correspondence, educational communities, craft networks: each of these can use digital and agentic infrastructure while maintaining a form of life whose temporality comes from elsewhere. They sustain themselves not by exiting the medium but by subordinating the medium’s temporality to a discipline the medium does not generate. A study group, religious community, or political organization may use agents to schedule meetings, translate materials, preserve archives, summarize discussions, and coordinate dispersed members without surrendering the temporality of study, ritual, obligation, or organizing. In such cases the agent assists a form of life whose binding comes from elsewhere. They remain vulnerable wherever their continuation depends on platform mechanisms — visibility, ranking, monetization, identity systems, payment rails, moderation, cloud access, agentic filtering — that import the dominant logic back into the practice. Under agentic conditions, the importation runs deeper, because the field within which the practice’s offerings appear to its participants is increasingly itself pre-composed. A religious community whose members encounter its existence through agent-curated information environments faces a more demanding subordinative work than one that did not.

The counter-tendencies show that the drift is not a mechanical destiny. They do not show that the drift can be reversed at the level of the medium by the practices alone. The conditions for their continuation are themselves infrastructural questions, and the practices’ resilience increasingly depends on whether the infrastructures they require can be governed in ways that respect temporalities the dominant logic does not share.

Resistance, if the word still means anything under agentic conditions, cannot consist in selecting different outputs from the same machinery of pre-composition. Nor can authenticity mean fidelity to preferences whose conditions of appearance have already been arranged elsewhere. What remains is not a program but a site of contestation: the systems that generate the field of choice before choice appears. The question is not whether the subject can still press approve. The question is whether the subject, or any institution still answerable to it, can participate in composing the world to which approval is given.