This paper examines the performance of the ZCS Michigan-style classifier system in multi-agent environments. Using an abstract multi-agent model the effects of varying aspects of the performance, reinforcement and discovery components are examined. It is shown that small modifications to the basic ZCS architecture can improve its performance in environments with significant inter-agent dependence. Further, it is suggested that classifier systems have characteristics which make them more suitable to such non-stationary problem domains in comparison to other forms of reinforcement learning. Results from the initial use of ZCS as an adaptive economic trading agent within an artificial double-auction market are then presented, with the findings from the abstract model shown to improve the efficiency of the traders and hence the overall market.