This paper addresses the architectural foundations of dynamic workflows in distributed multi-agent systems (MAS) integrated in Grid context. The purpose is to design an architecture at the same time taking into consideration tasks dependencies among agents, adaptation with respect to historic lessons learnt from past behaviour (memory) and the autonomous decisions when an unpredicted event occurs. In order to do this, given one ontology, called AGIO, which describes Agent-Grid Integration, we propose a workflow based on MAS with a complementary decision-making aid using Markov Logic Networks (MLN).