A three-dimensional typology of agency for advanced AI systems

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.20041 Date read: 2026-09-02 Connected to: L-008, L-011, L-012 Kind: content Escalation: store-only Escalation rationale:

What this is

A typological framework paper developing categories of AI agency across three dimensions (likely: intentionality, autonomy, and locus of control or similar), grounded in philosophy, ethics, and legal theory. The work is normative-descriptive: it classifies forms of agency rather than empirically investigating how agency emerges, breaks down, or shifts under optimization pressure or scaling.

What I took from it

The paper appears to address a real gap in conceptual clarity — distinguishing types of agency instantiated by frontier systems rather than asking only whether a system "has" agency. This could inform L-011 (Causal Detachment as Stable Equilibrium) and L-012 (Intervention-Layer Displacement) by providing vocabulary for how systems maintain operational functionality while becoming causally opaque or logically detached from their original intent.

However, the framework is typological and static, not mechanistic or dynamic. It does not investigate how agency types shift under adoption pressure, metric capture, or computable enforcement. It does not ask whether certain types of agency are more susceptible to ossification, proxy collapse, or causal detachment under scaling. It does not examine whether legibility constraints force systems into particular agency profiles, or whether those profiles are stable or phase-transition points.

The work is competent conceptual scaffolding but does not present a sustained empirical or theoretical argument about laws governing how these agency types behave under stress, competition, or formalization pressure.

Research connections

L-008: Typologies of agency might predict which optimization profiles become legible under computable enforcement, but the paper does not develop this dynamically.
L-011: The framework could describe which causal configurations constitute "functional" equilibria, but does not explain why those equilibria emerge or persist.
L-012: A typology could classify where the optimization locus lands in different agency architectures, but does not explain the displacement mechanism itself.
seed-062 (Formalization Opacity Collapse): Weak signal: as agency is formalized into typological categories, does the lived heterogeneity of actual systems become invisible? Undeveloped.

Seed

Seed title: none

Seed type: —

Seed text: —

Store recommendation: File under "AI agency conceptual frameworks." Flag for review if future papers begin to show empirically that systems migrate between agency types under adoption or metric pressure, or that certain types are inherently more prone to causal detachment or proxy capture. The typology may then become an induction asset.