Multi-Agent AI Systems Are Organizations

Multi-agent AI systems are described, in computer science, in terms such as communication protocols, reward functions and reasoning architectures. The authors argue this framing is incomplete and turn to an established framework: a century of organization science identifies universal problems that any multi-agent goal-oriented system must solve. These problems are substrate-neutral, and so they apply to multi-agent AI systems just as much as they do to human organizations. The authors show that the failures of multi-agent systems cataloged in an existing empirical taxonomy of 1,642 execution traces can be sorted almost perfectly into the universal problems of organizing. The problems are universal; the solutions that are feasible given agent properties and contextual conditions remain to be discovered. The authors discuss the implications of viewing multi-agent AI systems as organizations for systems comprising artificial agents as well as mixed human-AI members.