Swarms are often considered within the context of a large group of coordinated agents. Here, we consider a different social problem: Is it possible to identify and predict the behaviour of a coordinated swarm of agents within a larger swarm of unknown agents? Recent neurocognitive research indicates that properties of social agency may be captured within mechanisms for language. Accordingly, we considered new information-theoretic dynamical system models with which the behaviour is described using small language models. In this new modelling paradigm, systems are characterised in terms of short-term probabilistic micro-events forming self-emergent, synthetic languages. This approach creates the possibility for novel language-based signal interpretation, capturing social activities. In this frontiers style piece, we propose a model that incorporates this framework using independent component analysis as an approach to separate coordinated agent behaviour. Some experiments are presented, which indicate the potential feasibility of this methodology.