HPC systems face new challenges from coding agents running jobs faster
Towards Efficient HPC Systems for Agents: Challenges and Opportunities
Distributed, Parallel, and Cluster Computing
Summary
High-performance computing (HPC) systems were built for people running jobs, but now software agents are submitting many more jobs and using a large share of resources. These agents work much faster than humans and try many small changes to reach their goals, which stresses the system tools like schedulers and storage. The paper finds that current HPC designs don’t fit these agents well and suggests treating agents as special users that understand the system. The authors point out challenges and opportunities in how computing, storage, memory, and safety should evolve for these agents.
What this means in practice
- •For hpc system administrators: Design new scheduling and resource policies that recognize software agents as special tenants to better manage their rapid and metadata-heavy job submissions.
- •For cloud platform engineers: Develop storage and compute architectures optimized for trial-and-error workflows typical of coding agents, improving performance and resource use.
A position paper. It proposes an approach and reports no results.
Authors
Yunjia Zheng, Bintang Dwi Marthen, Zachary Pan, Minghao Li, Raminder Singh, Manasvita Joshi, Minlan Yu, Juncheng Yang
Abstract
Coding agents have become real users of high-performance computing (HPC) systems, yet today's HPC abstractions, interfaces, and policies remain designed for human-driven workflows. In our measurement, users running coding agents are only 19.5% of the observed population, but account for 55.8% of job submissions, 29.1% of CPU core-hours, and 42.7% of GPU-hours. Agents are not simply faster humans. They issue commands at 20.8x the human rate, decompose work into fine-grained explore-modify-execute loops, and pursue open-ended goals through trial-and-error campaigns that continue through nights and weekends. These behaviors strain abstractions built for human timescales, surfacing as control-plane pressure on the scheduler, metadata-intensive I/O on bandwidth-provisioned filesystems, repeated rediscovery of what earlier sessions already learned, and new prompt-injection and policy-enforcement surfaces. Neither banning agents nor treating them as ordinary users is sustainable. We instead argue for co-design, that facilities should treat agents as first-class principals where agents become facility-aware tenants. We outline the resulting challenges and opportunities in compute, storage, agent memory, and safety.