Reinforcement Learning for Dynamic Workflow Optimization in CI/CD Pipelines

Continuous Integration and Deployment (CI/CD) pipelines are core to modern software delivery, but their static workflows can be inefficient. This paper proposes a reinforcement learning (RL) approach to optimize CI/CD pipeline workflows dynamically. We model the pipeline as a Markov Decision Process and train an RL agent to make runtime decisions (e.g., selecting test scope) that maximize throughput while minimizing testing overhead. A simulated CI/CD environment with configurable build, test, and deploy stages is developed to evaluate the approach. Experimental results show that the RL-optimized pipeline achieves up to a 30 % improvement in throughput and about a 25 % reduction in test execution overhead compared to a static baseline. The agent learns to skip or abbreviate certain tests when appropriate, accelerating delivery without significantly increasing the risk of undetected failures. This work demonstrates the potential of RL to adapt DevOps workflows for greater efficiency, providing novel insights into intelligent pipeline automation.