Abstract
As organisations deploy Multi-Agent Systems (MAS) to automate complex business processes, they risk digitising a classic organisational failure by diffusing responsibility. We use complex adaptive systems (CAS) as a guiding theoretical lense to examine MAS. We provide an illustrative scenario of a manufacturer’s multi agent system to demonstrate that individual agent compliance is insufficient to prevent emergent systemic harms. We suggest principles for the design of good governance in complex systems of AI agents. We propose a research agenda for a Systemic Attribution Framework that provides governance mechanisms to evaluate multi-agent systems as composite entities. This ensures that granular compliance cannot shield systemic risk.