This paper presents a multi-agent simulation of the production of step-level public goods in social networks. In previous public goods experimental research the design of the sequence ordering of decisions have been limited because of the necessity of simplicity taking priority over realism, which means they never accurately reproduce the social structure that constrains the available information. Multi-agent simulation can help us to overcome this limitation. In our model, agents are placed in 230 different networks and each networks' success rates are analyzed. We find that some network attributes -density and global degree centrality and heterogeneity-, some initial parameters of the strategic situation -the provision point- and some agents' attributes -beliefs about the probability that others will cooperate-, all have a significant impact on the success rate. Our paper is the first approach to an explanation for the scalar variant of production of public goods in a network using computational simulation methodology, and it outlines three main findings. (1) A less demanding collective effort level does not entail more success: the effort should neither be as high as to discourage others, nor so low as to be let to others. (2) More informed individuals do not always produce a better social outcome: a certain degree of ignorance about other agents' previous decisions and their probability of cooperating are socially useful as long as it can lead to contributions that would not have occurred otherwise. (3) Dense horizontal groups are more likely to succeed in the production of step-level public goods: social ties provide information about the relevance of each agent's individual contribution. This simulation demonstrates the explanatory power of the structural properties of a social system because agents with the same decision algorithm produce different outcomes depending on the properties of their social network.