Uploaded October 2026 | Updated November 2026, 3 hours ago
Modern network engineers have more options than ever for testing changes before they reach production: virtual network operating systems, containerized labs, topology-as-code, CI/CD pipelines, and increasingly, network digital twins. But these approaches don’t all answer the same questions.
Over the past year, I built an internal network emulation platform designed to create realistic, multi-vendor network environments on demand. The goal was straightforward: make it easy to reproduce networks, test changes, inject failures, and validate behavior before touching production. Building it exposed an important distinction between emulating a network and modeling one.
This session explores that distinction through practical examples and lessons learned. We’ll look at where network emulation works exceptionally well, where its fidelity begins to break down, and why simply running more realistic virtual devices doesn’t necessarily produce a better model of production behavior.
We’ll then examine how behavioral modeling and digital twins approach the problem differently, particularly for questions such as reachability, policy changes, failure impact, and predicting the consequences of configuration changes.
Rather than arguing that one approach replaces the other, this talk will show how emulation, automated validation, and network modeling can complement each other—and what I learned trying to build a system that uses all three.
Craig Johnson:
Craig Johnson is a Principal Architect at Forward Networks, where he focuses on cloud strategy, network architecture, automation, and emerging technologies.
With more than 25 years of experience in networking and infrastructure, Craig has worked across enterprise networking, service providers, cloud platforms, and network security. His work spans both traditional network engineering and modern infrastructure, with a particular focus on making complex environments easier to understand, automate, and operate.
Craig is an active contributor to the networking community, with interests in network automation, digital twins, open-source infrastructure, and the practical application of AI to network operations. He regularly speaks at industry events and enjoys building tools and platforms that solve real engineering problems.
Modern network engineers have more options than ever for testing changes before they reach production: virtual network operating systems, containerized labs, topology-as-code, CI/CD pipelines, and increasingly, network digital twins. But these approaches don’t all answer the same questions.
Over the past year, I built an internal network emulation platform designed to create realistic, multi-vendor network environments on demand. The goal was straightforward: make it easy to reproduce networks, test changes, inject failures, and validate behavior before touching production. Building it exposed an important distinction between emulating a network and modeling one.
This session explores that distinction through practical examples and lessons learned. We’ll look at where network emulation works exceptionally well, where its fidelity begins to break down, and why simply running more realistic virtual devices doesn’t necessarily produce a better model of production behavior.
We’ll then examine how behavioral modeling and digital twins approach the problem differently, particularly for questions such as reachability, policy changes, failure impact, and predicting the consequences of configuration changes.
Rather than arguing that one approach replaces the other, this talk will show how emulation, automated validation, and network modeling can complement each other—and what I learned trying to build a system that uses all three.
Craig Johnson:
Craig Johnson is a Principal Architect at Forward Networks, where he focuses on cloud strategy, network architecture, automation, and emerging technologies.
With more than 25 years of experience in networking and infrastructure, Craig has worked across enterprise networking, service providers, cloud platforms, and network security. His work spans both traditional network engineering and modern infrastructure, with a particular focus on making complex environments easier to understand, automate, and operate.
Craig is an active contributor to the networking community, with interests in network automation, digital twins, open-source infrastructure, and the practical application of AI to network operations. He regularly speaks at industry events and enjoys building tools and platforms that solve real engineering problems.





![Keynote: Why Do We Have Self-Driving Cars, But Not Autonomous Networks?
Self-driving cars were imagined as early as 1918 [1], accelerated by the DARPA Grand Challenge era (2004–2007) [2], and are now beginning to operate at commercial scale.
Self-driving networks were framed as an industry ambition in Kireeti Kompella’s 2017 NANOG talk [3]. But in networking, “driving” has come to mean several different things: configuration automation, network control, or automated operations.
The industry has made significant progress in configuration automation and network control, including traffic engineering and VPN control. Yet most network operations still live in an alert → ticket → operator loop, assisted by runbooks and, increasingly, copilots and agents. This talk focuses on autonomous network operations, the missing piece for network autonomy.
So, nearly ten years on, why are self-driving cars becoming reality at scale, while autonomous network operations is still an industry ambition?
The answer isn’t that networks are somehow harder than cars. The issue is more fundamental. Networking is approaching autonomous operations as an observability and workflow problem, building on the systems we already have. That is the natural first step when applying AI to network operations. But self-driving cars have shown that autonomy is not simply another AI capability added to existing systems. It requires a paradigm shift in architecture, signals, models and control.
In this keynote, we apply the control-system approach used in self-driving cars and robotics to autonomous network operations, assess what is missing today, show where autonomy can be deployed safely first, and describe a realistic path for the industry to move from assisted operations towards autonomous control.
[1] Apex.AI — “The Driverless Car Era Began 100 Years Ago”
https://www.apex.ai/post/the-history-of-automated-driving-part-one-the-driverless-car-era-began-100-years-ago
[2] DARPA — “The DARPA Grand Challenge: 10 Years Later”
https://www.darpa.mil/news/2014/grand-challenge-ten-years-later
[3] Kireeti Kompella (NANOG, June 2017) — “The Self-Driving Network™: How to Realize It”
https://www.nanog.org/sites/default/files/1_Kompella_The_Networking_Grand_Challenge.pdf
John Evans:
John Evans works on the development of AI-native systems at the intersection of network architecture, telemetry, control and operations, with the goal of making large-scale network infrastructure more autonomous.
He has worked in large-scale network infrastructure for over two decades, including network architecture, traffic engineering, planning, telemetry, service assurance and operational control. Most recently, he was a Principal Network Engineer for Network AI at AWS, working across network planning, engineering and operations. Before AWS, he was a Distinguished Engineer at Cisco, focused on service provider architecture and strategy, and held technical leadership roles at Cariden and BT.
Across those roles, John has seen network operations evolve from alert and ticket pipelines toward increasingly automated systems. His current work focuses on what is required to move beyond incremental automation toward autonomous network operations.
John has represented AWS and Cisco at NANOG and the IETF, and currently co-authors IETF OPSAWG work on packet discard reporting and classification. He is a Fellow of the IET, CCIE #4125, a named inventor on multiple networking patents, and the author or co-author of publications including IP/MPLS QoS for Multiservice Networks and MPLS Next Steps. Keynote: Why Do We Have Self-Driving Cars, But Not Autonomous Networks?](https://i.ytimg.com/vi/Jf_2t1GF9mA/mqdefault.jpg)




