Enactive Cognitive Architectures for LLMs: A Homeostatic Approach to Functional Teleology and Algorithmic Individuation

Large Language Models (LLMs) currently exhibit high linguistic proficiency but lack the temporal continuity and intrinsic goal-directedness required for autonomous agency. Standard agentic frameworks typically rely on stateless prompting chains, which often results in "identity drift" and a lack of coherent long-horizon planning. This paper introduces the Homeostatic-Enactive Cognitive Architecture (H-ECA), a closedloop control system that regulates LLM output through internal state vectors (Competence, Connection, and Necessity). Unlike reactive architectures, H-ECA implements a biological-inspired separation of concerns: a Homeostatic Regulation Layer (HRL) for drive calculation, an Executive Strategy Module (ESM) for tactical planning, and a Self-Referential Episodic Store (SRES) for identity persistence. We present a longitudinal comparative analysis of two agent instances (α and β) initialized from an identical baseline. Experimental telemetry reveals that structural coupling with distinct interlocutors drives the system toward divergent behavioral attractors-a process we term Algorithmic Individuation. Furthermore, forensic analysis of execution logs demonstrates Access Consciousness, defined here as the computational capacity of the agent to monitor and accurately report its own latent decision-making logic. These results suggest that functional agency emerges not from model scaling, but from strict homeostatic error minimization protocols.