Agency Is Not Necessary — Systemic Optimization

— The generalization

$$\forall e \in \mathcal{E} : \exists W_e \quad [\text{agency is not required}]$$

Every entity that exists in a system has a welfare state. Agency is a special case.

— The ontological error

Why agents are not the right unit

Standard optimization — and standard AI alignment — takes the agent as the fundamental unit. But agents are a subset of a broader category: entities. Rivers, ecosystems, future generations, microbiomes — all are affected by decisions but have no agency.

The consequence: a framework built on agents cannot formally represent the welfare of non-agentic entities. Their exclusion is not deliberate — it is structural. The framework has no slot for them, so they never enter the objective function, which means they can never generate a corrective signal from within the system.

— The generalization

From agent to entity

Define entity $E$ as: any element of a system whose welfare state $W_e$ can be formally represented and for which decisions produce measurable welfare changes $\Delta W_e$.

$$\mathcal{E} = \{e : \exists W_e,\; \exists \Delta W_e(a)\}$$

— The entity set: all elements with a representable welfare state and measurable welfare change under actions.

The agent is a special case of entity: an entity that also has the capacity to take actions and observe outcomes. This does not make agents more important — it makes them a subset. Building a framework around the subset and ignoring the superset is the ontological error that 186 years of economic theory has reproduced.

— Theorem 3.1 (Observational Closure)

Observational Closure Theorem

Let $S$ be a system with operative entity set $E_{op}(S)$. For any entity $e \notin E_{op}(S)$:

(1) $\Delta W_e$ generates no signal in $S$'s objective function;
(2) $S$ cannot detect its own exclusion of $e$ from within;
(3) Therefore, verification of inclusion requires an external perspective.

Corollary: External audit of AI systems is not an ethical preference — it is a mathematical necessity. A system cannot verify its own alignment because any entity excluded from its operative set is invisible to its own objective function.

— Systemic optimization

The corrected objective function

$$F_1 = \sum_{e \in \mathcal{E}} \Delta W_e(a) \quad \text{subject to phenomenon constraints}$$

— The systemic objective: aggregate welfare change across all entities in the system, constrained by physical reality.

Three properties distinguish systemic optimization from standard optimization:

The entity set is open — any affected entity can be included. The boundary of $\mathcal{E}$ is empirical, not definitional.
Welfare functions need not be agent-declared — they can be observed, constructed from data, or built participatorily. Preference revelation is not required.
The constraint set is physical, not social — resources, thermodynamics, time. Budget constraints and institutional constraints are inputs, not axioms.

— Applications

What changes when entities replace agents

AI systems

Training with $F_1$ instead of individual reward means the system learns to maximize welfare of all affected entities from the start — not via RLHF constraints layered on top of a misaligned objective. The correction is at the level of the loss function, not the guardrail.

Policy

Cost-benefit analysis with entities instead of agents includes ecosystem tipping points, future generation welfare, and communities without political representation. The scope of the analysis is determined by who is affected — not by who has a vote or a preference declaration.

Law

Corporate personhood becomes a special case of entity representation. Rivers, ecosystems, and future generations can be formally represented with $W_e$ — which is already happening in Ecuador, New Zealand, and India. The formal framework provides the mathematical grounding for what those legal systems are already doing pragmatically.

Paper connections

Depends on

Sustains