A multi-agent partially observable Markov decision process (MPOMDP) is a\nmodeling paradigm used for high-level planning of heterogeneous autonomous\nagents subject to uncertainty and partial observation. Despite their modeling\nefficiency, MPOMDPs have not received significant attention in safety-critical\nsettings. In this paper, we use barrier functions to design policies for\nMPOMDPs that ensure safety. Notably, our method does not rely on discretization\nof the belief space, or finite memory. To this end, we formulate sufficient and\nnecessary conditions for the safety of a given set based on discrete-time\nbarrier functions (DTBFs) and we demonstrate that our formulation also allows\nfor Boolean compositions of DTBFs for representing more complicated safe sets.\nWe show that the proposed method can be implemented online by a sequence of\none-step greedy algorithms as a standalone safe controller or as a\nsafety-filter given a nominal planning policy. We illustrate the efficiency of\nthe proposed methodology based on DTBFs using a high-fidelity simulation of\nheterogeneous robots.\n