The ability to flexibly time movements is crucial for behavior. The motor cortex, basal ganglia, and thalamus form a multiregional neural network, known as the cortico-basal ganglia thalamocortical (CBGTC) loop, which controls various aspects of motor behaviors, including the timing of actions [1]. For example, in mice performing a delayed licking task, average activity in the striatum and cortex ramp up to a threshold for executing a lick [2]. However, the distinct contributions from different cell types and regions within the CBGTC loop remain largely unknown due to the similarity in their ramping dynamics and a lack of biologically constrained models that investigate the modular computations at hand. To overcome these limitations, we develop a circuit-constrained multi-regional recurrent neural network (mRNN) with cell types that replicate the CBGTC loop during the delayed licking task. Specifically, we model the direct and indirect pathways of the basal ganglia that target the substantia nigra pars reticulata (SNr) and disinhibit the thalamus, subsequently influencing the cortex while preserving Dale’s law. We train the model using gradient descent to replicate ramping dynamics in the cortical output region, viewed as the dominant signal in this task [3], and investigate the underlying solution learned by the network. To do so, we reverse-engineer the network and discover line attractor dynamics implemented within the subcortical thalamostriatal loop that properly times the cortical ramp, in agreement with integrator hypotheses of the striatum [4]. Silencing different regions in the mRNN model leads to concrete predictions on the resulting implications for the rest of the circuit, which we validate on optogenetic silencing experiments of the anterior lateral motor cortex (ALM) and striatum during the same task [5]. Our biologically plausible methodology elucidates the computational underpinnings performed by the circuit, and the circuit level connections that make this computation possible.