Platform teams considering AI agents in infrastructure workflows need to understand how policy gates, reviewer capacity, and recovery controls interact before production integration. This session presents a rule-based agent-based modeling (ABM) prototype with editable Bank AI integration workflow steps, AI agents, human reviewers, and system services. Attendees will see how to vary agent risk assumptions and compare data filters, policy checks, approvals, independent review, monitoring, and fallback using paired, seeded experiments. They will learn to interpret residual harm, detections, unresolved work, wait times, and relative cost—and to separate synthetic assumptions from operational evidence. Designed for platform engineers, SREs, and researchers, the talk offers a reproducible way to explore mitigation trade-offs and a starting point for extending the model to Kubernetes and GitOps delivery.