AI Agents Now Build Network 'Flight Simulators' on Their Own

WHAT HAPPENED: A team of researchers (Sudipta Acharya, Petar Djukic, Burak Kantarci) built something they call the 'NDT Factory.' It's a group of AI agents (software that can plan and take actions on its own) that write code to build a 'digital twin' of a network — essentially a working replica, like a flight simulator for phone and internet systems. Instead of engineers hand-coding simulation logic, the AI reads a plain description of what the network should do and produces the working simulator itself.

WHY IT MATTERS: Right now, when a phone company wants to change how it handles traffic, it either tests on live customers — risking dropped calls and outages — or pays engineers weeks to build custom simulation tools. This system could shrink that from weeks to hours. The team tested it on a common real-world task: deciding whether to accept or reject a new phone call when the network is busy (known as Call Admission Control, basically a bouncer at the door). Across 300 test scenarios, the AI-built simulator agreed with a hand-built reference version 99.3% of the time, accepted 90% of calls, and correctly explained every rejection. It compiled and passed tests 100% of the time.

THE CATCH: This is a research paper, tested on a single, fairly narrow case study — one type of network decision, not a whole telecom network. It hasn't been deployed at scale by a real carrier, and 99.3% agreement still means roughly 2 in 300 decisions differed from the human-built version. In networks serving millions of calls, small disagreements can matter. There's also the bigger question of who checks the AI's work when it's writing the tools we use to check the network.

THE BOTTOM LINE: The pattern here — AI agents generating the simulation tools that verify AI-managed systems — is likely to spread across telecom, cloud computing, and eventually any industry that runs on big infrastructure. For everyday people, the payoff is fewer outages and faster service improvements. For engineers who build network test systems by hand, it's a hint that part of that job is being automated. The paper was accepted to the IEEE Globecom conference in 2026.