This paper analyses the effects of manipulating the cognitive complexity of L2 oral tasks on language production. It specifically focuses on self-repairs, which are taken as a measure of accuracy since they denote both attention to form and an attempt at being accurate. By means of a repeated measures design, 42 lower-intermediate students were asked to perform three different tasks types (a narrative, and instruction-giving task, and a decision-making task) for which two degrees of cognitive complexity were established. The narrative task was manipulated along +/- Here-and-Now, an instruction-giving task manipulated along +/- elements, and the decision-making task which is manipulated along +/- reasoning demands. Repeated measures ANOVAs are used for the calculation of differences between degrees of complexity and among task types. One-way ANOVA are used to detect potential differences between low-proficiency and high-proficiency participants. Results show an overall effect of Task Complexity on self-repairs behavior across task types, with different behaviors existing among the three task types. No differences are found between the self-repair behavior between low and high proficiency groups. Results are discussed in the light of theories of cognition and L2 performance (Robinson 2001a, 2001b, 2003, 2005, 2007), L1 and L2 language production models (Levelt 1989, 1993; Kormos 2000, 2006), and attention during L2 performance (Skehan 1998; Robinson, 2002). (Contains 5 tables and 7 notes.)