With the exponential growth of global transaction volumes and the increasing sophistication of money laundering techniques, it has become more challenging for financial institutions to detect money laundering activities. There is an urgent need for more effective data mining models and more automated anti-money laundering (AML) technologies. This paper presents the construction of an AML intelligent agent, AMLAgent, which integrates large language models, graph neural networks, and traditional machine learning technologies. The entire process is centered around large models, utilizing multi-agent collaboration and multi-model integration to automatically accomplish tasks such as data processing, money laundering user identification, and report generation. This addresses the issue of a lack of interconnectivity among various models and the reliance on manual labor for data processing and report analysis. Experimental results indicate that our proposed AMLAgent has improved the F1-score in detecting money laundering users by 7.13%, and the generated reports are highly readable, offering good interpretability throughout the detection process.