In this paper we identify challenges that confront the largescale multi-agent system (LMAS) designer, and claim that these challenges can be successfully addressed by agent-based software engineering (ABSE), which we consider to be distinct from object-oriented software engineering for multi-agent systems (OOSE for MAS) in its consideration of agent goal, role, context and attitude as first class objects. We show how we have discovered these principles through our experiences in developing the RETSINA multi-agent system, in implementing specific test applications, and in the derivation of three distinct architectures that help guide and describe the designs of our systems: the individual agent architecture, the functional architecture, and the infrastructure architecture.