We propose a decentralized solution for high-level multi-agent task planning problems in environments considering communication network failure. In particular, we consider that robots can only sense each other and communicate within a limited radius, yet, they may need to collaborate to accomplish their tasks. These tasks are given in Metric Interval Temporal Logic (MITL), which is capable to capture complex task specifications involving explicit time constraints. To substitute for the lacking communication networks, we deploy an agile robot (e.g., drones) to transfer information between the heavy-duty robots while executing tasks. We propose an algorithm to decompose each MITL formula that is assigned to the corresponding heavy-duty robot into an independent task of that robot and an independent request for others. The agile robot systematically pursues heavy-duty robots to exchange requests. The heavy-duty robots use formal methods-based algorithms to compute path plans satisfying the independent promises and the received requests. While the robots' plan computation is fully decentralized, the satisfaction of all tasks is guaranteed (if such plans are found). The proposed solution can be applied to practical applications where the communication network fails or is restricted, such as post-catastrophe search and rescue and wildlife surveillance.