Though some research has focused on agent-direct affective processes, none has examined its impact on multi-agent learning environments and on the detection, modeling and fostering of self-regulated learning processes. 38 participants interacted with MetaTutor, an intelligent, multi-agent hypermedia-learning environment, to learn about the human circulatory system. The log files, containing information about their overall performance, and self-report measures, assessing emotions and impressions towards agents obtained from their interactions with MetaTutor were used to assess the relationship between subjective agent-directed emotions, SRL processes and overall learning gains. Results indicate that agent-directed emotions were not significantly related to SRL strategy use, negative agent-directed emotions were significantly related to negative learning gains and negative agent-directed emotions for two specific agents (representative of two SRL pillars) were related to negative learning gains. Implications for the design of multi-agent systems and the role of emotions during human-agent interactions and their relation to learning are discussed.