This article presents a new approach to the evolution of controllers for autonomous agents. We propose the evolution of a connectionist structure where each node has an associated program, evolved using genetic programming. We call this structure a Genetically Programmed Network and use it to successfully evolve control systems with very different architectures, by making small restrictions to the evolutionary process. Experimental results of applying this method to evolve neural networks, distributed programs and rule-based systems capable of solving a common benchmark problem, the Ant Problem, are presented. Comparison with other known genetic programming based approaches, shows that our method requires less effort to find a solution.