Agent-breakage: A measurement framework for autonomous Kubernetes operations

A new paper from arXiv (2605.23058) introduces agent-breakage, a closed-loop measurement framework for autonomous Kubernetes operations agents. The authors argue that empirical claims about such agents are largely unfalsifiable due to lack of controlled baselines, selection bias, absence of pre-registered decision matrices, and small sample sizes. The framework injects faults into a target Kubernetes cluster, observes agent responses, and scores them on four axes against ground truth, accumulating outcome-labeled tuples. It distinguishes framework error from reasoning error and supports off-condition testing. The work aims to provide a verification substrate analogous to code agents' testing environments.