This paper examines recent proposals for task-based approaches to instruction. It reviews relevant research, before going on to examine a number of potential problems with task-based teaching, such as a potential focus away from form and towards lexis. It reviews recent developments in cognitive psychology which support a dual-mode perspective for language processing, and then proposes the goals of accuracy, complexity-restructuring, and fluency as the most relevant for task-based instruction. In the final section the paper proposes a framework for the implementation of task-based instruction which draws upon relevant theory and research, and which organises the methods by which such instruction could be put into practice in such a way as to minimise problems, and maximise the probability that all three above goals can be achieved.