Many problems can only be specified statistically and machine learning techniques may be the most effective ways for their solution. I've evaluated various ML technologies (e.g. genetic algorithm, genetic programming, Bayesian nets) for use in a variety of different classes of problems (e.g. game playing, travelling salesperson problem, user interfaces).
Given the evolution of the Web, autonomous agents will happen. What we need is a theory that explains and predicts the behavior of communities of interacting agents.