Scaling AI Safely with Bounded Multi-Agent Systems

The frontier of AI research is dominated by highly capable, resource-hungry, general-purpose, monolithic agents, which pose significant safety risks that are difficult to mitigate. We argue for a greater focus on studying and building specialised, narrow AI agents, whose capabilities are intentionally limited with safety and efficiency in mind. Moreover, we argue that scaling the deployment of AI in society can be achieved more safely and efficiently through multi-agent systems comprised of such agents. We refer to such systems as bounded multi-agent systems (BMAS). We present a case for BMAS and discuss the outstanding technical challenges and novel safety risks they pose.