Abstract #
AI agents are changing systems research on two fronts at once. Frameworks that automate the design-and-evaluate loop already produce competitive heuristics with little human effort, raising fears that the field’s core craft is being automated away. This talk argues for a more optimistic reading. Unlike other domains, where AI threatens to absorb nearly all of the work, agents in systems compress implementation and testing, leaving ideation and real-world validation as the parts of research that people do. At the same time, agents are a new workload with new requirements, the kind of shift that has always produced the field’s best problems. On using agents for systems, the talk presents TuxBot, which uses LLM reasoning to tune Linux continuously on live machines and outperforms state-of-the-art tuners by 145% on average. On building new systems for agents, it shows how speculative actions cut the latency of sequential agent workflows, and why the safe exploration they require necessitates a rethinking of the system stack.
Bio #
Kostis Kaffes is an assistant professor in the Department of Computer Science at Columbia University after obtaining his MSc and PhD from Stanford University in 2018 and 2022, respectively. He is broadly interested in the intersection of computer systems, artificial intelligence, and cloud infrastructure, aiming to enable both robust agentic AI deployments and next-generation cloud performance. His work has been recognized with Sony, Google, and Meta research awards. He is a founding member of Columbia’s Data, Agents, and Processes Lab (DAPLab). Prior to Columbia, he spent a year at Google SRG.