Triplet loss, a widely utilized function in deep gait recognition, optimizes embedding spaces by pulling features of the same identity while pushing those of different identities. However, we found that the Euclidean distance, which is frequently adopted in triplet loss, inherently limits the optimization process in the later stage of training. Through gradient decomposition analysis, a key limitation is recognized, i.e. as samples approach positive anchors, the dominance of their radial gradient components diminishes, hindering convergence. To address this issue, we propose the Angle-guided Triplet Loss (ATL), which incorporates a geometrically inspired rotation function that re-calibrates gradient directions based on the between-sample angles. ATL effectively enhances radial guidance from positive samples, thereby accelerating convergence and improving the discriminative power. Extensive experiments demonstrate the superiority of ATL, as it consistently outperforms state-of-the-art methods across various model scales and loss functions on challenging in-the-wild datasets.