Abstract-The paper focuses on the problem how a community of distributed agents may autonomously invent and coordinate lexicons and grammars. Although our earlier experiments have shown that a communication system can indeed emerge in a socio-cultural dynamics, it relies on the control of complexity by the experimenter, so that agents first acquire words, then simple constructions, and then more complex ones. This paper addresses the question how agents could themselves regulate the complexity both of the mechanisms they bring to bear to the language task and on the semantic complexity of what they want to express. We make use of the autotelic principle, coming from psychology. It requires monitoring challenge and skill (based on actual performance) and maintaining a 'flow' regime balancing the two. We show in computational experiments that the autotelic principle is able to explain autonomous scaffolding towards greater complexity in the emergence of language.