Emerging networked applications, such as AI services, 5G/6G RAN, Internet of Things (IoT), autonomous driving, and extended reality (XR), demand highly performant, reliable, and secure network infrastructures. In response to these demands, many efforts envision an ambitious “self-driving workflow” for network management that makes real-time control decisions to meet the needs of applications. Just like self-driving cars, self-driving networks should be able to “see,” “analyze,” and “control“ diverse network behaviors to achieve high performance, reliability, and security going forward. In this talk, I will discuss the challenges of letting networks run themselves and how our research enables a viable pathway to tackle these challenges. First, I will discuss how we can provide accurate and real-time observability of diverse application-level metrics on heterogeneous network platforms (e.g., programmable switches and end hosts). Second, with real-time observability, I will describe a case study on when the networks can run themselves to defend against distributed denial-of-service attacks. Finally, I will chart paths to designing future AI tools for network operations (AI for NetOps).