The proposed project, BAYESMART, introduces a new class of smart control agents that combine the rigor of Model Predictive Control (MPC) with the adaptive power of Bayesian optimization. Traditional embedded controllers are static and inflexible, while most AI agents for smart environments rely on opaque, resource-hungry models. BAYESMART bridges this gap by transforming MPC from a fixed optimizer into a self-tuning intelligent agent that can perceive its performance, reason under uncertainty, and adapt in real time. This ambition moves beyond the state of the art by embedding Bayesian routines directly into resource-constrained platforms (GPUs, FPGAs, and MCUs). Experimental proof of concept (paper accepted to the most prestigious conference in process control: 64th IEEE Conference on Decision and Control) shows that resulting smart agents are capable of autonomously balancing stability, efficiency, and actuator wear, delivering at least 2× faster responses, 30% higher energy efficiency, and robust accuracy preservation compared to standard implementations. The project advances technology from TRL 3 to TRL 5, validated in representative smart environments such as industrial process control or autonomous sensing platforms. By releasing the developed libraries as open source, BAYESMART contributes directly to the European edge AI ecosystem, aligning with AI Act requirements, open standards, and sustainable innovation goals. In this way, MPC becomes more than a controller — it becomes the core of a Bayesian-driven smart agent that senses, decides, and acts locally, placing Europe at the forefront of trustworthy and adaptive edge intelligence.