Biological intelligences often consist of many different, complex systems working together, rather than a single mechanism capable of solving every problem. It may be that artificial intelligence requires a composition of multiple, diverse systems; perhaps more than any one individual or one research group can create or even understand. This dissertation presents an architecture for hosting and efficiently running user-submitted programs (minds) intended to solve particular problems (worlds), and to facilitate the assembly of these problem-solving minds into larger scale systems of hybrid minds which query an existing set of minds (which we call subminds) for their suggested actions. These subminds may be hosted on the same machine or remotely, across the Internet. They may have been written by many different authors, with each program intended either as a complete solution to the problem, or with an eye toward its re-use in a modular fashion. Even if a program is written and intended as a complete, independent solution to the problem, that need not preclude its inclusion in a larger, hierarchical hybrid program. In other words, the architecture presented forms a platform for building hybrid artificial intelligence systems using other people's code. The dissertation also presents and evaluates a method for automatically combining minds (which may have been written by different authors) for use in a hybrid mind, by performing a statistical analysis of the observed performance of each mind program in a collection of minds submitted by third parties competing to solve two different types of problem.