This paper describes a multi-agent model of double-auction market in which simulations are led with artificial agents. The study of market can be focused on the study of information processing for the agents, and this means that one has to make assumptions about the cognitive use that the agents do of these information. For a few years, experiments have been used to study auctions. The resulting data have more re-cently been used to make hypothesis about learning and have occasionally been translated in algorithms. We propose here simulations that are organised on the same model as experiments, as a succession of auctions session where each agent is given a good at the beginning of the session and has to exchange it before the end. Agents are either seller or buyer and make bids and asks along the time, that can be accepted by the others and lead to transactions.  Our main result is the fact that we actually obtain convergence although agents have no knowledge of others' limit prices and only interact through a completely impersonal market, which correspond to experimental data.