NeTS: Exploring the Limits of Agentic AI in PhD-level Networked Systems Design

Modern networked systems, including the cloud platforms and wireless networks that society depends on every day, have grown enormously complex. They span vast scale, varied hardware, many layers of software, and constantly shifting demands, all while needing to be fast, reliable, and affordable. Designing and tuning these systems is slow and demanding work, and the people who do it are increasingly overloaded. As complexity grows faster than human effort can keep pace, innovation slows and the risk of costly failures rises. Recent advances in artificial intelligence (AI), particularly large language models (LLMs), raise the possibility that AI agents could help engineers do this work. This project investigates a foundational question: whether AI systems that can experiment, reason, and form hypotheses can take part in the kind of expert design work that today requires highly trained researchers and engineers. The goal is not to replace human judgment but to understand where AI can genuinely help, where it falls short, and why. The work emphasizes designs that humans can understand, clear statements of what each design is meant to do, and human oversight. If successful, the project could reduce the cost, time, and human expertise needed to build the infrastructure that underpins commerce, communication, and public services. The project also produces benchmarks, datasets, and evaluation tools for the research community. It aims to discover which parts of this expertise can be aided by AI and which remain inherently human, helping to shape how future engineers are trained and engaging students in computing. The project studies networked systems design as a structured reasoning process rooted in the scientific method that spans the discovery of decision rules, the formation of useful abstractions, and the design of architectural decisions. The research is organized into progressively harder challenges. First, it examines the discovery of better heuristics in settings where goals are measurable and evaluation is benchmarkable, using concrete case studies across networked systems. It studies how AI can discover such rules that remain dependable when conditions shift at runtime and that can adapt safely while in operation. Second, it explores whether AI can rediscover or propose the abstractions and architectural concepts that define influential systems research, a task that resists measurement by any single number. A central feature of the approach is grounding every AI-generated design in precise goal and in high-fidelity evaluation through simulation, emulation, and testbeds. A further contribution is a set of methods for judging AI-generated designs not only by performance but also by harder-to-measure qualities such as understandability, robustness, and conceptual clarity. To study abstraction discovery rigorously, the project develops a replay method that reconstructs the problem context of prior systems research and asks whether an AI system can independently arrive at comparable concepts. Throughout the work, human experts remain involved through hypothesis review, specification checks, and expert critique, so that the emphasis stays on comprehension and scientific rigor. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.