The paper describes a software architecture for autonomous excavation tasks. For this, it uses a behavior-based approach which allows adaptation to the changing environment and unexpected events. As part of the presented architecture, basic primitives are defined which solve spatially limited motions. These can be activated as and when required. It is possible to have multiple of them running in parallel. Normally, an excavation task can be split up into several different phases. During these phases, the primitives will be activated as needed. To be reusable, the primitives are designed in a very general way. To show the performance of the presented architecture, an example application is built up which can be used to dig a trench. This application is implemented on a system running on a real backhoe loader.