The purpose of this paper is to investigate how complexity, accuracy and fluency interact in early L2 development, when learners’ linguistic means are underdeveloped. Learners then resort to rote-learned formulaic sequences to complement their current grammar when it is unable to meet their communicative needs. The interplay between their nascent grammar and these formulaic sequences provides an interesting window onto how learners resolve tensions between complexity, accuracy, fluency, and communicative needs.Formulaic sequences are very common in early L2 productions and enable learners to communicate in spite of limited linguistic means so that they appear to be more advanced in the L2 than they actually are, in terms of complexity, accuracy and fluency. In spite of their prevalence in early productions, however, the role that these formulaic sequences play in L2 development remains unclear. Are they used as communicative crutches until learners’ grammatical competence enables them to generate these forms productively or do they contribute more directly to learners’ linguistic development? This paper reports on an empirical study tracing and analysing the development of such sequences over time in the early L2 productions of instructed learners of French. It is argued that these sequences gradually become unpacked during the acquisition process, and are used as models by learners in order to assist them in the construction of a productive L2 grammar. The tension between, on the one hand, complex, accurate and fluent formulaic sequences, but whose internal structure has not yet been analysed into its constituents in order to use them productively elsewhere, and on the other hand, an underdeveloped linguistic system which does not allow communicative needs to be met, seems to be driving the acquisition process forward in these learners. The linguistic status of these sequences seems to be that of single multimorphemic lexical units which have been assigned a semantic representation but are underspecified syntactically. Their intrinsic morphosyntactic complexity, therefore, and the fact that they can be accessed as single units rather than effortfully constructed from scratch in real time communication, gives the misleading impression of complex, accurate and fluent productions at a stage when learners’ productive grammars are in fact very simple, approximate and non-fluent.