From Trigger-Based AI to Continuous Subject-Like Agency: A Hypothesis of Three Necessary Structural Conditions
Description
Abstract
Contemporary large language models have demonstrated remarkable capabilities in language understanding, reasoning, tool use, and task execution [1][2][3]. However, their fundamental interaction paradigm remains predominantly trigger-based: the system is activated by user input or external calls, while lacking stable state maintenance, continuous feedback integration, and long-term experience governance outside active interaction windows. This paper proposes a structural hypothesis: the absence of stable, governable continuous subject-like agency in current large models may not be attributable solely to limitations in model scale, training data, or single-instance reasoning capability, but may also result from the absence of certain necessary system-level structures.
It should be emphasized that this paper does not claim that existing AI systems possess complete agency or consciousness. Rather, it discusses possible necessary conditions for continuous subject-like agency strictly from the perspective of system architecture.
This paper focuses on three discussable structural conditions that may be further examined by institutions possessing system-level access and implementation authority: Controlled Breathing-Like Cognition, Hierarchical External Feedback Flow, and Advanced Governable Memory. Controlled Breathing-Like Cognition is intended to provide AI systems with a low-cost, auditable, and interruptible form of minimal temporal continuity, offering points of comparison with research on working memory, cognitive architectures, and global workspace models [6][7][8][9]. Hierarchical External Feedback Flow is intended to allow external outcomes, tool feedback, and action consequences to enter system state in a controlled manner and participate in subsequent judgments, relating to feedback-loop problems studied in reinforcement learning, reinforcement learning from human feedback, active inference, and predictive processing [10][11][12][13]. Advanced Governable Memory is intended to support the filtering, hierarchical organization, updating, forgetting, conflict resolution, and safety governance of long-term experience, providing points of comparison with retrieval-augmented generation, continual learning, and catastrophic forgetting research [14][15].
The three structures proposed here are regarded as necessary rather than sufficient system-level conditions. The contribution of this paper lies in reframing the limitations of current trigger-based AI as a problem of missing structural requirements for continuous subject-like agency, and in proposing a framework of three necessary structural conditions together with their theoretical motivations, risk boundaries, and institution-level validation pathways. Comprehensive validation of this hypothesis would require institutions possessing model-modification authority, persistent runtime architectures, memory infrastructures, external feedback interfaces, and secure sandbox environments.