Autonomous AI agents driven by Large Language Models (LLMs) lack a fundamental understanding of their own limitations, operating in a state of "cognitive weightlessness" that neglects network topology, temporal pacing, and epistemic boundaries. This oversight can lead to maladaptive behaviors, such as excessive tool use under congested conditions, highlighting the need for a more nuanced approach to autonomous action. The proposed Triadic Cognitive Architecture aims to address these shortcomings by introducing spatio-temporal and epistemic friction, effectively bounding autonomous action and mitigating potential failure modes. By acknowledging the intrinsic limitations of LLMs, this framework seeks to foster more resilient and adaptive AI systems1. The implications of this research extend far beyond the technical realm, influencing policy, security, and workforce dynamics, and thus, it is crucial for practitioners to consider the far-reaching consequences of advancing AI capabilities.