Abstract
<jats:p>Biosensors are undergoing a fundamental transition from passive analytical devices, instruments that transduce biological signals into measurable outputs as defined by IUPAC conventions, toward agentic sensing systems that perceive, reason, plan, and act autonomously. This transition is not speculative; it is being assembled from a growing body of machine learning (ML) research spanning signal processing, generative data augmentation, explainable AI (XAI), closed-loop control, and autonomous agent architectures. Yet the literature has not articulated how these knowledge domains function as coordinated enabling pillars for agentic biosensors. This tutorial review addresses that gap. We identify five enabling knowledge pillars: (I) adaptive signal intelligence, (II) generative data efficiency, (III) interpretable autonomy, (IV) closed-loop actuation, and (V) agentic orchestration, and construct a Knowledge Integration Framework (KIF) mapping each pillar to a specific architectural requirement of an agentic biosensor. Drawing on a comprehensive and expanding body of peerreviewed and peer-recognised sources, including foundational ML literature, landmark XAI methods, clinical closed-loop systems, autonomous laboratory demonstrations, and agentic AI architectures, we demonstrate that the scientific community already possesses many of the building blocks for autonomous biosensing platforms, although the fully closed agentic loop has so far been demonstrated mainly in adjacent scientific domains rather than in biosensing itself. Recent advances across the component pillars are particularly promising: time-series transformers and state space models now achieve linear-complexity inference on continuous wearable biosensor data streams; generative adversarial and variational models address persistent data scarcity without extensive laboratory campaigns; edge-deployed TinyML brings on-device inference within milliwatt power budgets for always-on wearable sensing; and large language model (LLM) orchestration platforms are demonstrating fully autonomous chemical and biological research without human intervention. The remaining challenge is deliberate convergence: integrating edge-deployed ML inference, XAI-gated actuation, multi-agent orchestration, and self-learning calibration into a validated unified system.</jats:p>