With the extensive application of multi-task learning in complex business scenarios such as intelligent recommendation and autonomous driving, the handling mechanism for heterogeneous task priorities has emerged as a critical bottleneck constraining model performance. Existing approaches predominantly rely on empirical strategies of linear loss function combination, which suffer from dual deficiencies: First, The manual parameter-tuning paradigm for task weighting leads to combinatorial explosion in hyperparameter search space as task quantity increases; second, gradient competition among tasks induces negative transfer effects that significantly erode the performance boundaries of high-priority tasks. To address these challenges, this study proposes an innovative constrained optimization-based progressive multi-task learning method. By encoding task priority structures into inequality constraints, we formulate an optimization paradigm with strict priority guarantees. A constrained optimization framework is established through Lagrangian duality theory to ensure that the performance lower-bound constraints of high-priority tasks remain unaffected by secondary task optimization processes. Meanwhile, a progressive gradient projection algorithm enables dynamic adjustment of constraint spaces. Theoretically, we provide convergence guarantees through non-convex optimization theory. Experimental results on public datasets demonstrate that our method enhances the performance of secondary tasks while ensuring the stability of high-priority tasks, establishing a novel theoretical framework and technical pathway for multitask learning.