This paper presents a protocol that permits to automatically allocate tasks, in a distributed way, among a fleet of agents when communication is not permanently available. In cooperation settings when communication is available only during short periods, it is difficult to build joint policies of agents to collectively accomplish a mission defined by a set of tasks. The proposed approach aims to punctually coordinate the agents during "Rendezvous'' phases defined by the short periods when communication is available. This approach consists of a series of simultaneous auctions to coordinate individual policies computed in a distributed way from Markov decision processes oriented by several goals. These policies allow the agents to evaluate their own relevance in each task achievement and to communicate bids when possible. This approach is illustrated on multi-mobile-robot missions similar to distributed traveling salesmen problem. Experimental results (through simulation and on real robots) demonstrate that high-quality allocations are quickly computed.