Watching your own back: self-managing multi-agent systems

We describe a category of multi-agent applications that address the problem of managing widely distributed Multi-Agent Systems (MAS) by introspectively managing both themselves and other co-resident MAS applications. The result is an extremely fault-tolerant system that has well-defined survivability parameters without requiring deep coordination between the primary application and the management layer. We argue that such an integrated systems-management solution can be both effective and simpler than an out-of-band solution would allow, and can leverage all the flexibility of a MAS infrastructure. We describe and analyze an exemplar of such a self-managed system.