Agents operating in the real world have to deal with a constantly changing and only partially predictable environment and are nevertheless expected to choose reasonable actions quickly. This problem is addressed by a number of action-selection mechanisms. Behaviour networks as proposed by Maes are one such mechanism, which is quite popular. In general, it seems not possible to predict when behaviour networks are well-behaved. However, they perform quite well in the robotic soccer context. In this paper, we analyse the reason for this success by identifying conditions that make behaviour networks goal converging, i.e., force them to reach the goals regardless of the details of the action selection scheme. In terms of STRIPS domains one could talk of self-solving planning domains.