← back to paper
arxiv: 2604.25899 · 2 revisions
Pythia: Exploiting Workflow Predictability for Efficient Agent-Native LLM Serving