One current research direction in multi-agent systems is architectural decomposition of LLMs: rather than one monolithic model, train multiple specialized LLMs (one for math, one for intuition, one for creativity) and route tasks to appropriate models—which is itself a form of multi-agent work within a single LLM system.