Bimanual human activities inherently involve coordinated movements of the hands and body, which are also grounded in the surrounding visual context such as the objects being manipulated. Despite this tight coordination, existing benchmarks are not designed to study this synergy systematically: annotating either hand or body pose alone, or relying on marker-based capture that alters the visual appearance away from natural recordings. To bridge this gap, we introduce AssemblyHands-X, the first markerless 3D hand-body benchmark for understanding bimanual human activities. Building on the multi-view markerless videos of the Assembly101 dataset, we reconstruct temporally consistent SMPL-X hand-body motion for 2.8M images through a multi-view annotation pipeline. Using this dataset, we systematically analyze how the synergy among hand, body, and visual cues influences bimanual action recognition. Through experiments with multi-view annotated poses on AssemblyHands-X and single-view estimated poses on the large-scale Assembly101 dataset, we show that (1) combining 3D hand and body cues significantly improves verb recognition, and (2) modeling these pose cues jointly with video further boosts action recognition performance. These findings highlight the importance of modeling multimodal synergy for a holistic understanding of bimanual activities.