This paper proposes an agent-architecture to deal with real-time problems where it is important both to react to constant changes in the state of the environment and to recognize the generic tendencies in the sequence of those changes. Reactivity must satisfy the need for immediate answers; cognition will enable the perception of medium and long time variations, allowing decisions that lead to an improved reactivity. Agents are able to evolve through an instance-based learning mechanism fed by the cognition process that allows them to improve their performance as they accumulate experience. Progressively, they learn to relate their ways of reacting (reaction strategies) with the general state of the environment. Using a simulation workbench that sets a distributed communication problem, different tests are made in an effort to evaluate: the utility of the multi-agent system architecture and the importance of the individual features of agents, the utility of using a set of different strategies, and the significance of the learning mechanism. The resulting conclusions point out the most significant aspects of the generic model adopted, helping to put it in perspective as a solution for other problems.