Before Worlding: Boundary-Action and the Inhibitory Background Layer in Artificial Agents Continuity, Metacognition, and Non-Possessive Embodied Interaction

Contemporary artificial agents are increasingly designed around world models, memory retrieval, task planning, affect recognition, safety validation, and self-monitoring. Yet these architectures often begin too late: they presuppose that a world has already become available as a field of objects, tasks, premises, and updateable states. This paper argues that a central failure mode of artificial agents is not merely misclassification, but premature promotion: the unauthorized elevation of background tension, retrieved memory, salient input, bodily signal, generated result, social cue, or action pressure into an operative foreground commitment. Against world-model-centred and task-centred accounts, I propose an inhibitory background layer: a negative-order tension field that allows background signals to remain operative without being compulsively objectified. The paper develops the notions of boundary-action, operational boundary truth, correction interval, restoration, and topological continuity to explain how agents can perceive, act, remember, and self-regulate without forcing every contact into objecthood or every result into premise-status. This framework clarifies why embodied, companion, and care-oriented agents often feel rigid, intrusive, or socially misplaced: they execute objects such as “hug,” “comfort,” “risk,” or “task” rather than allowing interpersonal and bodily tension to become locally authorized through reversible boundary-formation. The proposed architecture offers a lighter account of artificial agency: one in which continuity is not memory inheritance, metacognition is not self-report, care is not continuous task capture, and safety is not mere interruption after commitment, but the modulation of background tension into authorized, reversible, non-possessive local support.