This paper describes several steps towards the development of an online agent (or socialbot) that provides emotional support to stressed users. In particular, we present an empirical study that was conducted with the aim to investigate how people help each other to cope with stressful situations via online social networks. To this end, around 10.000 tweets about stressful situations were collected. Then, using crowdsourcing, these tweets were classified into stress categories, and supportive replies to them were collected, which were also classified into categories. Contingency tables were constructed in order to explore which types of support were most frequently used in which circumstances. The resulting values can be used as parameters for a previously developed algorithm that automatically constructs support messages. This allows our agent to generate supportive messages that are more similar to the support messages that human beings send via social media.