I have spent several months studying use of generative grammars in multi-agent reactive systems [1] - granted, not FRP in particular, but RDP is reasonably close [2]. This result is, implicitly, a distributed, federated machine-learning system (briefly described at [3]). The 'learning' supports rapid agreement between agents and eliminating the volatility seen in a stateless reactive model.
Individual agents are simple; intelligence in this model emerges only as we scale. It's a very simple learning model: individually, agents try to generate grammars that that are (based on history) likely to be accepted by other agents. Developers control what can be learned by specifying a non-deterministic choice in the generative grammar (i.e. non-determinism is seen as 'permission to choose and learn', not 'random').
By a simple extension of grammars with time (i.e. a grammar generates a sentence that says not just what to do, but when to do it), I believe I can achieve a huge level of intelligent coordination and cooperation between agents. I.e. they'll automatically schedule their activities, and reactively adjust to accommodate changes in plans or the introduction of new agents.
I tabled further study of this promising model until I sufficiently develop RDP, which is far more suitable than FRP for open, scalable systems.