Team-level retrospectives are widely used in agile and lean software development, yet little is known about what is actually discussed during retrospectives or their outcomes. In this paper, we synthesise the outcomes of sprint retrospectives in a large, distributed, agile software development organisation. This longitudinal case study analyses data from 37 team-level retrospectives for almost 3 years. We report the outcomes of the retrospectives, their perceived importance for process improvement and relatVed action proposals. Most discussions were related to topics close to and controllable by the team. However, the discussions might suffer from participant bias, and in cases where they are not supported by hard evidence, they might not reflect reality, but rather the sometimes strong opinions of the participants. Some discussions were related to topics that could not be resolved at the team level due to their complexity. Certain topics recurred over a long period of time, either reflecting issues that can and have been solved previously, but that recur naturally as development proceeds, or reflecting waste since they cannot be resolved or improved on by the team due to a lack of controllability or their complexity. For example, the discussion on estimation accuracy did not reflect the true situation and improving the estimates was complicated. On the other hand, discussions on the high number of known bugs recurred despite effective improvements as development proceeded.