The Frame-Stability Problem in Decision-Theoretic Accounts of Agency
Abstract
Contemporary decision-theoretic, planning-theoretic, and control-based accounts of agency explain how agents optimise, plan, and respond to reasons over time. What they do not explain is why there must be a single agent to begin with, rather than a succession of locally rational decision processes. This article diagnoses a shared structural presupposition across these accounts: each assumes a stable evaluative standpoint in which preferences, utilities, commitments, and revisions are already unified across time. I argue that this assumption conceals a prior problem—the frame-stability problem—concerning how any system can remain a coherent locus of evaluation under diachronic uncertainty when the evaluative frame itself is generatively unstable under learning, feedback, and environmental change. Optimisation, planning, and control mechanisms coordinate action only conditional on such stability; they cannot by themselves secure the persistence of evaluative authority. I propose that agency arises as a structural solution to this pressure: a system counts as an agent when it must preserve itself as a unified evaluator across incompatible future trajectories. Unified arbitration is introduced as the architectural role through which evaluative authority is maintained over time.
Keywords
- agency
- decision theory
- continuity
- AI governance
- accountability
- infrastructure
- commitments