Real-world optimization of complex products (e.g., planes, jet engines) is a hard problem because of huge multi-objective and multi-constrained search-spaces, in which many variables are to be adjusted while each adjustment essentially impacts the whole system. Since the components of such systems are manufactured and output values are obtained with sensors, these systems are subject to imperfections and noise. Perfect digital twins are therefore impossible. Furthermore simulating with sufficient details is costly in resources, and the relevance of Population-based optimization approaches, where each individual is a whole solution to be evaluated, is severely put in question. We propose to tackle the problem with a Multi-Agent System (MAS) modeling and optimization approach that has two major strengths: 1) a natural representation where each agent is a variable of the problem and is perceiving and interacting through the real-world topology of the problem, 2) a cooperative solving process where the agents continuously adapt to feedback, that can be interacted with, can be observed, where the problem can be modified on-the-fly, that is able to directly control these variables on a real-world product while taking into account the specifics of the components. We illustrate and validate this approach in the Photonics domain, where a light beam has to follow a path through several optical components so as to be transformed, modulated, amplified, etc., at the end of which sensors give feedback on several metrics that are to be optimized. Robotic arms have to adjust the 6-axis positioning of the components and are controlled by the Adaptive MAS we developed.