Pythia: Exploiting Workflow Predictability for Efficient Agent-Native LLM Serving
Abstract
As LLM applications grow more complex, developers are increasingly adopting multi-agent architectures to decompose workflows into specialized, collaborative components, introducing structure that constrains agent behavior and exposes useful semantic predictability. Unlike traditional LLM serving, which operates under highly dynamic and uncertain conditions, this structured topology enables opportunities to reduce runtime uncertaintyyet existing systems fail to exploit it, treating agentic workloads as generic traffic and incurring significant inefficiencies. Our analysis of production traces from an agent-serving platform and an internal coding assistant reveals key bottlenecks, including low prefix cache hit rates, severe resource contention from long-context requests, and substantial queuing delays due to suboptimal scaling. To address these challenges, we propose Pythia, a multi-agent serving system that captures workflow semantics through a simple interface at the serving layer, unlocking new optimization opportunities and substantially improving throughput and job completion time over state-of-the-art baselines.
Cite
@article{arxiv.2604.25899,
title = {Pythia: Exploiting Workflow Predictability for Efficient Agent-Native LLM Serving},
author = {Shan Yu and Junyi Shu and Yuanjiang Ni and Kun Qian and Xue Li and Yang Wang and Jinyuan Zhang and Ziyi Xu and Shuo Yang and Lingjun Zhu and Ennan Zhai and Qingda Lu and Jiarong Xing and Youyou Lu and Xin Jin and Xuanzhe Liu and Harry Xu},
journal= {arXiv preprint arXiv:2604.25899},
year = {2026}
}