MapReduce is a design pattern for processing large datasets on a cluster. Its performances depend on some data skews and on the runtime environment. In order to tackle these problems, we propose an adaptive multiagent system. The agents interact during the data processing and the dynamic task allocation is the outcome of negotiations. These negotiations aim at improving the workload partition among the nodes within a cluster and so decrease the runtime of the whole process. Moreover, since the negotiations are iterative the system is responsive in case of node performance variations. In this paper, we show how, when a task is divisible, an agent may split it in order to negotiate its subtasks.