The development direction of transformer monitoring and diagnosis is remote and intelligent. But the existing systems have the following problems: lack of diagnostic knowledge and methods, poor flexibility etc. To solve these problems, the architecture of a remote monitoring and diagnostic system for transformers based on multi-agent techniques is presented. On the basis of ultra high frequency (UHF) partial discharge (PD), core earth current and dissolved gas analysis (DGA), the hierarchical decomposition model is established. The tasks, functions and structure involved in the collaboration agents are defined. Then, the fusion strategy from multi-experts conclusions is proposed based on diagnostic confidence and weights. In final, the system implementation is described. Some functions of the system have been put into service and worked well.