Goal-directed behaviour hinges on mental representations that encode instrumental relationships between actions and their outcomes. The present study investigated how people acquire such representations during joint actions whose outcomes depend on synchronized contributions of multiple co-actors. Adapting a well-established two-stage action-outcome learning task to a synchronous joint action context, we tested if co-acting individuals would link jointly produced tone outcomes to features of their individual action contributions only or to group-level features of their joint action contributions instead. While our results reveal limitations of associative action-outcome learning in joint action contexts, they show that action-outcome representations acquired in joint action contexts encode group-level relations between co-actors’ joint action contributions and affect subsequent construal of ambiguous task instructions. Our findings inform current debates on the limits and mechanistic nature of action-outcome learning and provide novel evidence that planning and control of joint action is supported by dedicated representational structures encoding group-level relations between own and others’ action contributions.