Simulating the Experienced Coach: An Orchestrated Multi-Agent Artificial Intelligence Architecture Decomposed by Physiological System for Continuous Athlete Monitoring

Athlete monitoring in elite sport requires the simultaneous interpretation of heterogeneous data streams spanning external load, internal load, neuromuscular status, sleep, nutrition, psychological well-being and injury history. An experienced coach performs this integration in parallel and continuously; most computational decision support systems, by contrast, compress this multidimensional state into a single prediction function. The present study addresses that mismatch. The

Aim: is to reframe athlete monitoring through a multi-agent architecture that mirrors the layered, parallel and continuously communicating regulatory structure of the human organism, to specify that architecture at a level that permits implementation, and to present it together with a testable validation protocol. The

Design: Science Research cycle was followed in four stages: derivation of

design requirements through an integrative synthesis of the athlete monitoring, physiological regulation, cybernetics and distributed artificial intelligence literatures; development of the architecture; scenario-based demonstration in a professional team sport context; and formulation of a pre-registrable validation protocol. The architecture, named HOMEOSTAT, decomposes the athlete into thirteen agents that map one-to-one onto physiological organ systems: from cardiovascular and respiratory to energy metabolism, from musculoskeletal to endocrine, from immune to psychoneuroendocrine. Each agent owns its metric streams end to end; it acquires, cleans, individualizes and interprets them. An orchestrator invokes these agents and aggregates their returns into six judgments corresponding to the questions a coach actually asks: loadability, recovery status, injury and illness risk, adaptation trajectory, competition readiness and fitness to train. The reflex interlock stands outside the orchestration loop and can halt it. Each agent is specified as a three-layer artificial intelligence component: a deterministic signal layer, a learned inference layer calibrated by conformal prediction, and an optional language reasoning layer that is forbidden from producing any numerical claim of its own. The decomposition principle is anatomical rather than architectural: agent boundaries follow organ system boundaries, because metrics cluster according to the system that produces them, domain expertise is organized in the same way, and a physiological system is the smallest unit about whose state a clinician can form a meaningful judgment. Orchestration was preferred over choreography, because accountability, deterministic replay and verifiability are preconditions in a domain deemed high risk under artificial intelligence regulation in the employment context. Three mechanisms are introduced. The modulator mechanism holds scalar signals that decay with a half-life; the orchestrator resolves them once per cycle and injects them into every agent invocation, thereby shifting decision thresholds multiplicatively across the whole system without issuing commands, and reproducing the effect of endocrine regulation while preserving determinism. The negative selection mechanism structurally overcomes the class imbalance that constrains existing injury prediction models, by learning the athlete's individual normal profile from abundant data instead of learning from rare injury events. The coach-model agent represents the expert's tacit knowledge through case-based reasoning and thereby renders it auditable. The Law of Requisite Variety demonstrates that a one-dimensional metric cannot regulate a multidimensional athlete state, and provides an information-theoretic foundation for the existing empirical critique of single-metric approaches. The contribution is not a validated system but a complete and extensible specification: agent contracts, model classes, confidence semantics, guardrail constraints, the training regime and the extension procedure are given at a level that permits implementation and reuse. The architecture has not been empirically validated, and no peer-reviewed empirical research on autonomous agent systems exists in the field; the priority is therefore not new architectural proposals but the rigorous testing of those already advanced.