We describe a novel approach for shop poor control combining a distributed multi-agent structure and computational intelligence techniques. Shop floor activities are controlled by a network of autonomous agents. Each agent makes its decision with a fuzzy algorithm evaluating all the alternative actions with multiple criteria based on real time measures of shop's conditions. A tuning mechanism of the decision algorithm allows agents to adapt themselves to the time varying operating conditions of the manufacturing system. The adaptation process follows a reinforcement learning schema. New agents are periodically created to replace the old ones according to the following strategy: the better the peformance of an agent in its life cycle, the higher the probability that new agents will inherit its decision rules. Preliminary experiments on a detailed simulation model of flexible assembling systems show the potentialities of the approach and suggest further improvements.