Granular excavation is the removal of solid, discrete particles from a structure composed of these objects. Efficiently predicting the stability of an excavation during particle removal is an unsolved and highly nonlinear problem, as the movement of each grain is coupled to its neighbors. Despite this, insects such as ants have evolved to be astonishingly proficient excavators, successfully removing grains such that their tunnels are stable. Currently, it is unclear how ants use their limited information about the environment to construct lasting tunnels. We attempt to unearth the ants’ tunneling algorithm by taking three-dimensional (3D) X-ray computed tomographic imaging (XRCT), in real time, of Pogonomyrmex ant tunnel construction. By capturing the location and shape of each grain in the domain, we characterize the relationship between particle properties and ant decision-making within an accurate, virtual recreation of the experiment. We discover that intergranular forces decrease significantly around ant tunnels due to arches forming within the soil. Due to this force relaxation, any grain the ants pick from the tunnel surface will likely be under low stress. Thus, ants avoid removing grains compressed under high forces without needing to be aware of the force network in the surrounding material. Even more, such arches shield tunnels from high forces, providing tunnel robustness. Finally, we observe that ants tend to dig piecewise linearly downward. These results are a step toward understanding granular tunnel stability in heterogeneous 3D systems. We expect that such findings may be leveraged for robotic excavation.