Accurate and automatic delineation of the left atrium (LA) is crucial for computer-aided diagnosis of atrial fibrillation-related diseases. However, effective model training typically requires a large amount of labeled data, which is time-consuming and labor-intensive. In this study, we propose a novel LA delineation framework. The region of LA is first detected using an actor-critic based deep reinforcement learning method with a shape-adaptive detection strategy using only box-level annotations, bypassing the need for voxel-level labeling. With the effectively detected LA, the impacts of class-imbalance and interference from surrounding tissues are significantly reduced. Subsequently, a semi-supervised segmentation scheme is coined to precisely delineate the contour of LA in 3D volume. The scheme integrates two independent networks with distinct structures, enabling implicit consistency regularization, capturing more spatial features, and avoiding the error accumulation present in current mainstream semi-supervised frameworks. Specifically, one network is combined with Transformer to capture latent spatial features, while the other network is based on pure CNN to capture local features. The difference prediction between these two sub-networks is exploited to mutually provide high-quality pseudo-labels and correct the cognitive bias. Experimental results on two public datasets demonstrate that our proposed strategy outperforms several state-of-the-art methods in terms of accuracy and clinical convenience.