Autonomous, software intensive systems increasingly combine data driven behavior with distributed cloud native infrastructure. This combination produces a reliability and security paradox: learning components must evolve rapidly, while infrastructure and governance controls must remain stable, auditable, and resilient under adversarial pressure. This paper proposes a unified autonomous systems architecture that integrates three historically separate threads: machine learning based defect prediction for proactive quality assurance, decision intelligence for policy aware lifecycle governance, and infrastructure security for end-to-end risk reduction. The architecture unifies telemetry, feature engineering, decision models, and security controls into a closed loop autonomy plane that continuously senses system health, predicts defect and incident risk, recommends actions under explicit economic and compliance constraints, and executes changes through controlled pipelines. We define a reference model, data contracts, and control loops spanning code, build, deployment, and runtime phases. A structured evaluation methodology is presented to measure predictive utility, operational benefit, and security posture improvements without conflating offline model accuracy with online impact. The result is an actionable blueprint for engineering autonomous systems that are measurable, governable, and defensible.