Helium: Workflow-Aware LLM Serving Achieves 1.56x Speedup Over State-of-the-Art Agent Systems

arXiv 2603.16104·high signal

Agentic workflows consist of interdependent LLM calls with overlapping prompts and intermediate results that existing systems like vLLM ignore, treating each call in isolation. Helium (arXiv:2603.16104) models agentic workloads as query plans with LLM invocations as first-class operators, adding proactive KV cache pre-warming for static prefixes and a cost-based, cache-aware scheduling algorithm using a templated radix tree. Achieves up to 1.56x speedup over state-of-the-art agent serving systems with no model changes required.