Ethereum is a blockchain platform that supports smart contracts. Smart contracts are pieces of code that perform general-purpose computations. For instance, smart contracts have been used to implement crowdfunding initiatives that raised a total of US$ 6.2 billion from January to June of 2018. In this paper, we conduct an exploratory study of smart contracts. Differently from prior studies that focused on particular aspects of a subset of smart contracts, our goal is to have a broader understanding of all contracts that are currently deployed in Ethereum. In particular, we elucidate how frequently used the contracts are (activity level), what they do (category), and how complex they are (source code complexity). To conduct this study, we mined and cross-linked data from four sources: Ethereum dataset on the Google BigData platform, Etherscan, State of the DApps, and CoinMarketCap. Our study period runs from July 2015 (inception of Ethereum) until September 2018. With regards to activity level, we notice that it is concentrated on a very small subset of the contracts. More specifically, only 0.05% of the smart contracts are the target of 80% of the transactions that are sent to contracts. New solutions to cope with Ethereum’s limited scalability should take such an activity imbalance into consideration. With regards to categories, we highlight that the new and widely advertised rich programming model of smart contracts is currently being used to develop very simple applications that tend to be token-centric (e.g., ICOs, Crowdsales, etc). Finally, with regards to code complexity, we observe that the source code of high-activity verified contracts is small, with at most 211 instructions in 80% of the cases. These contracts also commonly include at least two subcontracts and libraries in their source code. The comment ratio of these contracts is also significantly higher than that of GitHub top-starred projects written in Java, C++, and C#. Hence, the source code of high-activity verified smart contracts exhibit particular complexity characteristics compared to other popular programming languages. Further studies are necessary to uncover the actual reasons behind such differences. Finally, based on our findings, we propose an open research agenda to drive and foster future studies in the area.