Feedback-Driven Autonomy: Learning Safe Automation Boundaries In Large-Scale Decision Systems

Abstract

While automated decision systems are increasingly deployed at unbounded scales, autonomy is necessary for performance and availability. Autonomy is controlled only through thresholds, manual tuning, or infrequent human review, leading to automation boundaries that drift out of alignment with real-world risk once systems are in production. While previous work considers when to enable autonomy at runtime, there is relatively little work on how autonomy should adapt after deployment, during ongoing production operation. Feedback-driven autonomy is a system-level model where the boundaries of autonomy are adjusted in production not only based on prediction accuracy but also on observed operational performance. The control plane also manages an autonomy learning loop, which consumes asynchronous and aggregated feedback (e.g., rollback costs, incident correlation, human interventions) from live systems and adjusts the autonomy level of the system over time. Rather than learning an agent's decision-making like a reinforcement learning algorithm would do, the governance layer learns how autonomous the system should be in its current operational context. Examples in recommendation, offer vending, and notification systems inform the exploration of feedback-driven autonomy as a strategy for safe, scalable, and attribution-tolerant automation in high-throughput production environments. Uncovering autonomy as a learned property of the system provides a practical focus for designing adaptive, resilient, automated decision systems that can respond to a dynamic, real-world operational context after deployment.