Crowdsourcing is a phenomenon that is transforming the way in which companies use the internet to gather ideas, solve complex cognitive problems and develop high-quality collections of data and knowledge (Wikipedia), or even software (Linux), through self-organising agents. Many recent studies have highlighted factors and small sets of parameters that play a role when a large crowd interacts with an organisation. However, no comprehensive simulation has yet been developed that integrates all these parameters and generates predictive power. This work describes, based on a MAS, the development of a simulator for large crowds called 'CrowdSim', which collectively solve problems in a crowdsourcing scenario. The simulator enables sensitivity analyses of multiple parameters as well as simulations of interactions within complex networks of irrational agents. Furthermore, the simulator's modular and extensible structure allows users to improve accuracy and predictive power as researchers incorporate new empirical findings.