Controlling complex systems, such as combustion engines, imposes to deal with high dynamics, non-linearity and multiple interdependencies. To handle these difficulties we can either build analytic models of the process to control, or enable the controller to learn how the process behaves. Tuning an engine control unit (ECU) is a complex task that demands several months of work. It requires a lot of tests, as the optimization problem is non-linear. Efforts are made by researchers and engineers to improve the development methods, and find quicker ways to perform the calibration. Adaptive Multi-Agent Systems (AMAS) are able to learn and adapt themselves to their environment thanks to the cooperative self organization of their agents. A change in the organization of the agents results in a change of the emergent function. Thus we assume that AMAS are a good alternative for complex systems control. In this paper, we describe a multi-agent control system that was used to perform the automatic calibration of an ECU. Indeed, the problem of calibration is very similar to the one of control: finding the adequate values for a system to perform optimally.