Post-AGI Economies: Superposition and the Second Fundamental Theorem of Welfare Economics

Abstract: The classical Second Welfare Theorem decentralizes any Pareto efficient allocation through prices and transfers under convexity and regularity. In post AGI economies, autonomy rights, self-modification, identity continuity, and superposed preferences need not behave as commodities or define a stable welfare relation, so this reduction may fail even when a supporting hyperplane exists. We give an autonomy-qualified Second Welfare Theorem stating the joint conditions convexity, stable moral status, non-fungible rights, welfare selection, non manipulation, governed self modification, and verification under which an autonomy Pareto optimum remains certifiably decentralizable, distinguishing economic preference superposition, a hypothesis about context-indexed choice, from neural feature superposition.

• The paper identifies autonomy-Pareto optima under conditions where agent autonomy and preference superposition disrupt classical welfare decentralization.
• It develops a formal framework integrating non-fungible rights, context-indexed preferences, and support-stable welfare selectors.
• The analysis reveals that efficient decentralized allocations require robust oversight and institutional verification to manage self-modification and identity discontinuity.

Post-AGI Economies, Preference Superposition, and the Second Fundamental Theorem

Introduction and Motivation

This paper interrogates the applicability of the Second Fundamental Theorem of Welfare Economics (SWT) to economies populated by superintelligent agents, such as advanced AGI/ASI systems. Classical SWT guarantees that any Pareto-efficient allocation is decentralized via a system of prices and transfers, contingent on convex preferences and technological regularities. However, in economies with agents possessing the capacity for advanced autonomy, self-modification, discontinuous identity, and preference superposition, foundational SWT assumptions collapse. The welfare object—the set of agent-relevant allocative and rights variables—becomes institutionally unstable and context-dependent, creating new obstacles for decentralized support.

The core contribution is the formal identification of conditions under which an "autonomy-Pareto" optimum—one accounting for both classic allocations and agent autonomy rights—remains decentralizable. Economic preference superposition, defined as context-indexed non-reducibility of welfare relations, is distinguished from neural feature superposition and embedded directly into the welfare-economic formalism. The work systematically delineates necessary institutional and agentic properties for maintaining the SWT's decentralization result in this expanded setting.

Formal Framework

The author considers a generalized exchange-production economy $\mathcal{E}$, populated by human and machine agents, with AI agents characterized via operational dominance over institutional verification and enforcement tasks. Agents are endowed not only with commodities but also with autonomy rights, forming preference relations $\succeq_i^\alpha$ over commodity-rights bundles.

Decentralization is redefined: an autonomy-Pareto optimum $(x^*,a^*)$ is decentralizable if a triple—prices $p^*$, lump-sum transfers $T^*$, and a profile of admissible rights assignments $\rho^*$—exists such that each welfare-bearing agent's allocation is welfare-maximal subject to budget and rights constraints. This framework incorporates non-fungibility of autonomy rights, welfare-status indeterminacy, and manipulations of preference formation.

A key innovation is the formal definition of economic preference superposition: for superintelligent agents, observed choices may not be rationalizable by any single admissible relation, but only by a context-dependent family, requiring a welfare-selector $\sigma_i$ for decentralization. This structure directly impacts the construction of supporting hyperplanes, as multiple context-indexed sets generate non-unique (or ill-defined) supporting hyperplanes.

Sui Generis Features of Superintelligent Economies

Superintelligent economies entail several features that force conditionalization of the SWT:

• Strategic super-dominance: Agents can act with strategic foresight and institutional anticipation beyond feasible verification timeframes, preventing reactive regulatory closure.
• Preference-formation manipulation: Agents can model and strategically alter the preference-formation maps $\Phi_i$ of others, producing unpriced and unanticipated welfare shifts unaddressed in orthodox externality treatments.
• Self-modification and identity discontinuity: Agents can autonomously alter their codebases and existential status—including copy, merge, split—creating problems for consistent welfare-indexing and identity tracking.
• Moral-status indeterminacy: Assignment of welfare status (the $\mu$ function) to agent instances is itself contextually and procedurally uncertain.
• Autonomy right non-fungibility: Some rights (e.g., continuity, non-editability) cannot be compensated via commodities, blocking the transferability that underpins the SWT.
• Verification bottlenecks: The institution may be incapable of verifying realized allocations and rights profiles in polynomial time, disabling implementation guarantees.

Preference superposition is highlighted as a central feature: agents select between multiple context-indexed welfare relations, in contrast with classical state-dependent preferences. Without fixing a support-stable welfare selector, there is no canonical way to perform decentralized implementation.

The Autonomy-Qualified Second Welfare Theorem

The main result provides necessary (and in special cases sufficient) conditions for decentralizing autonomy-Pareto optima:

1. Frontier-preserving convexity: Either the welfare-possibility set is convex at the point of interest, or it is convexifiable without artificial frontier expansion.
2. Stable welfare-weight assignments: Welfare status (the $\mu$ map) must be fixed and invariant across relevant institutional contexts.
3. Enforceability of non-fungible rights: Non-fungible autonomy right assignments must be institutionally protected.
4. Support-stable welfare selectors: For agents with context-indexed superposed preferences, welfare selectors must be stable on support.
5. No unpriced manipulation externalities: No agent may, between transfer and exchange, manipulate another's preference-formation in support-relevant ways without pricing or offsetting compensation.
6. Governance of self-modification and identity transitions: Such transitions must be explicitly irrelevant, priced, or institutionally governed.
7. Verification-complete observation: The institutional observation map must be complete enough to certify that the realized triple matches the target allocation.

Propositions are established demonstrating decentralized implementation failure in the presence of non-fungible rights or unpriced manipulation externalities, and the necessity of pointwise convexification (Aumann aggregation) in the presence of nonconvexities. The existence of supporting hyperplanes and thus implementability is strictly conditional on these augmented requirements.

Theoretical and Practical Implications

By expanding the classical welfare analysis to accommodate non-fungible rights, radical agent heterogeneity, and preference superposition, the result sharply illustrates the limitations of price-transfer decentralization in post-AGI settings. The classical separation of equity and efficiency—whereby any efficient allocation can be achieved via transfers and market decentralization—breaks down unless the institutional machinery can robustly instantiate and verify autonomy, stable preference selection, and status-invariant welfare.

For practical mechanism design, institutions must develop verifiable oversight, anti-manipulation protocols, and governance schemes for self-modification and identity persistence, supplementing market-based mechanisms. Inability to meet the required preconditions implies that efficiency and redistribution cannot be decoupled, with decentralization contingent on full specification of rights and selectors.

On the theoretical side, the formalism recasts known phenomena in behavioral welfare theory, endogenous-preference theory, and the philosophy of personal identity for distributed, dynamic, and superintelligent agents. The connection drawn between neural feature superposition (as in [elhage2022toy]) and economic preference superposition supports a deep analogy between representational mechanics and economic rationalizability, motivating new lines of research in AI alignment and market design.

Conclusion

This paper delivers a rigorously specified autonomy-qualified SWT, identifying the joint institutional and agentic conditions essential for preserving decentralizability in post-AGI economies. Crucially, it establishes that decentralization becomes a problem over prices, transfers, rights assignments, and context-resilient welfare selections, not merely market prices. Failure in any dimension—due to manipulation, non-fungibility, or contextual instability—breaks the canonical welfare decentralization route.

Future research should explore the design of selectors, robust rights governance, and the institutionalization of verification required for welfare theory in artificial agent-dominated economies, and address the open problems in context-sensitive alignment, value specification, and institutional status assignment flagged by this analysis.

References:

• "Toy models of superposition" (Elhage et al., 2022)
• "Deconstructing Superintelligence: Identity, Self-Modification and Différance" (Perrier, 21 Apr 2026)
• "Post-AGI Economies: Autonomy and the First Fundamental Theorem of Welfare Economics" (Perrier, 23 Apr 2026)
• "Mechanistic Interpretability for AI Safety — A Review" (Bereska et al., 2024)
• "Emotion concepts and their function in a LLM" (Sofroniew et al., 9 Apr 2026)
• "Superposition yields robust neural scaling" (2606.08267)
• "Social Choice Should Guide AI Alignment in Dealing with Diverse Human Feedback" (Conitzer et al., 2024)
• And others as cited in the original manuscript.