Recent technological developments in robotics and artificial intelligence may enable the deployment of robots in many aspects of our lives. As the complexity of robotic platforms increases, deliberation algorithms need to be improved, in particular to handle an increasing number of agents, to manage complex goals and tasks, and to evolve in more open environments in which unforeseen events should be dealt with autonomously. Among the deliberation functionalities deployed to provide the best level of autonomy: planning, monitoring, learning, observing, we focus here on acting. We present OMPAS, a refinement-based acting engine that executes high-level tasks by refining them into a set of lowerlevel tasks and commands. OMPAS uses a custom Lisp dialect (SOMPAS) to define the behavior of the robotic agent. SOMPAS provides primitives to handle concurrency and resources, and allows the synthesis of planning models, thanks to the restricted core language, and the explicit identification of acting decisions. The engine has been extended to deploy a continuous planning module, using the synthesized models, to look ahead and guide the decisions of the acting system in order to e.g., avoid deadlock, or optimize the completion of several parallel tasks. We provide an evaluation of the overall approach on the control of a fleet of robots in a simulated logistic platform.