This paper is motivated by the task of modeling routing decisions with sequential forwarding interactions in multi-agent networks with topology uncertainties, e.g., agents' mobility traces with uncertain speed and direction, links to someone unknown in stranger social networks, both making their interactions come across at uncertain time and location. Since most routing designs assume that agents' behaviors be regular or with known probability distribution and envision the future of topology as stable/predictable (namely limited uncertainty), these approaches may suffer the difficulty dealing with the networks with high uncertainties. The proposed research aims to provide an effective solution for message routing among agents in such networks. Specifically, we introduce a new principle of causal entropy force in multi-agent networks for routing with high uncertainties, provide a new thinking way of routing, backward thinking, and build connections between individual intelligence, topology uncertainties, and message routing through path entropy in phase space. The experiment results with real dataset (30K taxies) indicate that the proposed method could achieve 83% message delivery rate, compared with 20%-25% of traditional approaches, and generally achieve much less latency compared with typical methods.