The optimization of complex industrial systems represents a class of difficult problems, due to their embodiment in the physical world, and whose search spaces are disrupted, non-linear and potentially vast. Their parametrization relies on the combination of many variables, each change generally impacting the whole system. Mathematical approaches are limited by the fact that the models are too coarse or nonexistent, and by the imprecision of the measurements and the machining of components of such systems. The action cost of the system calls into question population-based heuristics and swarm intelligence where each individual must be tested. We propose a multi-agent approach allowing a global / black-box modeling of the system in terms of input variables and objectives, as well as an agnostic and continuous adaptive optimization, based on sensor feedbacks from the running system.