Control and Safety Boundaries for Agentic AI in Cloud-Native Distributed Systems

The use of Agentic AI in cloud-native distributed software systems, to automate decisions, routing, optimization, and operations management, is on the rise. Agentic systems behave independently, adapt at runtime, and influence downstream system state in their own way without human intervention. These capabilities increase the responsiveness and scale. However, they also bring new types of system-level risk that are not addressed by the existing reliability and faulttolerance models. Autonomous decision loops provisioned by Agentic AI may continue to operate while providing unstable and economically destructive outcomes. This paper argues that Agentic AI must be governed through explicit control and safety boundaries at the systems architecture level. We propose binding designs for controlling Agentic AI autonomy in cloudnative distributed systems. We also emphasize control, observability, and deterministic intervention as mandatory infrastructure concerns.