Many companies are using MapReduce applications to process very large amounts of data. Static optimization of such applications is complex because they are based on user-defined operations, called map and reduce, which prevents some algebraic optimization. In order to optimize the task allocation, several systems collect data from previous runs and predict the performance doing job profiling. However they are not effective during the learning phase, or when a new type of job or data set appears. In this paper, we present an adaptive multi-agent system for large data sets analysis with MapReduce. We do not preprocess data and we adopt a dynamic approach, where the reducer agents interact during the job. In order to decrease the workload of the most loaded reducer-and so the execution time-we propose a task reallocation based on negotiation.