Agent-based modelling for historical societies runs offline and produces analytical results. Real-time multi-agent systems for virtual worlds come from game AI and lack historical parameterisation. This paper proposes the combination: a multi-agent architecture for historically grounded virtual worlds in which simulation runs in real time, a human flâneur inhabits the world, and narrative emerges from agent dynamics without scripting or game mechanics. The architecture is organised as a conceptual cube: nine simulation domains (economics, weather, agriculture, social structure, technology, conflict, ecology, culture, health) operating at three scales (micro, meso, macro), coordinated by two orchestrators (world state consistency and narrative coherence). NPC behaviour is produced by a layered model: historical constraints define the action space, social norms structure expectations, personal goals drive decisions, and compact memory/relationship state provides continuity. Historical parameterisation enters through a Regional Retrieval-Augmented Generation system that supplies era-specific and location-specific data to each domain agent. Narrative emerges through causal cascades — drought to scarcity to conflict to adaptation — identified by composable story-sifting patterns. Scaling strategies include cognitive level-of-detail for NPCs, crowd-level statistical models beyond the observation boundary, and cloud/edge computation partitioning. The architecture is a design proposal informed by established ABM foundations, game AI precedents, and emergent narrative theory; empirical validation through prototype implementation is identified as the primary next step.
James Otto Danenberg (Thu,) studied this question.