Precise software reliability management under dynamic and non-stationary operational environments remains a major challenge in modern software engineering. Although operational-profile-based and machine-learning-based models provide accurate reliability predictions, they lack adaptive decisionmaking capabilities for runtime optimization. To address this limitation, this paper proposes DRL-LCSROPE, a deep reinforcement learning–based extension of the LCSR-OPE and ML-ER-OPE frameworks for closed-loop, adaptive reliability and effort optimization. The proposed methodology integrates dynamic operational profile learning, level-wise composite reliability attribute generation, and joint effortreliability prediction within a Markov Decision Process. A policy-based deep reinforcement learning agent, utilizing Proximal Policy Optimization, learns optimal control strategies by observing system states that comprise composite reliability, cumulative effort, failure intensity, and operational profile entropy. The framework is evaluated on four NASA software defect datasets, JM1, KC1, KC2, and PC1, using repeated stratified 10-fold cross-validation. Experimental results demonstrate 10–15% reliability improvement, 1218% reduction in testing effort, and statistically significant gains over LCSR-OPE, ML-ER-OPE, and static operational-profile baselines at the 95% confidence level.