Microsoft describes layered controls for an AI agent handling network incidents

Microsoft engineer Arun Malik described on October 13 how the company built an AI agent to investigate incidents in Azure’s production network. The agent can collect telemetry and device state and classify a likely fault, but the model is not automatically authorized to make production changes.

The post describes access bounded by the invoking person’s identity, individually authorized tool interfaces, human approval before production changes, and converting repeated successful reasoning paths into deterministic playbooks. It is Microsoft’s account of its own engineering work, not an independent performance or safety evaluation.