Habits -- automatic behavioral patterns formed through repetition -- are essential for daily functioning, but can also lead to inflexible behavior. While crucial for understanding both adaptive and maladaptive decision-making, studying habits' computational and neural mechanisms has been challenging due to limited laboratory experiments demonstrating overtraining-induced inflexibility. We developed a novel task with features designed to encourage participants to engage goal-directed (GD) control between trials (interleaving extensively- and minimally-practiced contexts), then naturally release control within trials (hierarchical multi-step trial structure and opportunities to self-correct). Results showed that overtrained participants displayed stronger biases toward behaviors learned in extensively-practiced contexts, evidenced by higher Habit Index values at early response times. This effect decreased at later response times, suggesting participants could override habitual impulses with GD control. Our computational model, characterizing behavior as a mixture of reinforcement-learned policies, reproduced observed behavioral patterns, suggesting that habits can be viewed as goal-directed deployment of overtrained policies.