Many behaviors demand a balance between exploitation, which requires reliability, and exploration, which involves flexibility. In cortical circuits, much work has emphasized how attractor dynamics generate reliable behavior through the convergence of a network to specialized activity patterns for specific decisions and actions1. Other work has emphasized how cortical activity exhibits variability across trials of a behavior, which could in principle support flexibility in actions2. Here, through multi-area calcium imaging and recurrent neural network models, we propose that cortical activity for navigational decisions separates into distinct attractor regions allowing for behavioral reliability, while the dynamics within each attractor is disordered, permitting controlled variability for trial-to-trial flexibility. Based on simultaneous calcium imaging from neurons in V1, posterior parietal cortex, retrosplenial cortex, and M2, we developed a novel environment-interacting recurrent neural network model constrained by data through imitation learning. The trained model is a dynamical system that autonomously generates single-trial neural activity and locomotion behaviors matching the data recorded from mice performing a navigation task. Perturbations of the model revealed distinct attractor states for each navigation decision, and analysis of the connectivity in the model consistently showed competition between the decisions through opponent inhibition. Surprisingly, however, within each attractor neural activity did not return to their original trajectories following small perturbations and instead drifted to new trajectories, which is a hallmark of disordered dynamics. Such disordered dynamics were mediated by connections creating competition between subgroups of neurons tuned to different navigational trajectories even for the same decision. A background of net inhibitory connections stabilized the disordered dynamics from running away. Together, our work proposes a critical role for disordered attractors in decision-making, and their underlying circuit mechanisms, where attractor dynamics provide reliability and disordered dynamics generate variability for flexibility.