In spite of many success stories in various domains, genetic algorithms and genetic programming still suffer from some significant pitfalls. Those evolved programs often lack important properties such as robustness, comprehensibility, transparency, modifiability and usability of domain knowledge easily available. We attempt to resolve these problems, at least in evolving high-level behaviours, by adopting a technique of conditions-and-behaviours originally used for minimizing the learning space in reinforcement learning. We experimentally validate the approach on a foraging task