Janus: Edge Content Negotiation for Agent-Oriented Web QoS
Year: 2026
Type: conference-paper
Abstract: LLM agents increasingly retrieve web content through tool calls, yet current web delivery paths remain optimized for human browsing rather than machine reasoning. A typical page carries layout, scripts, image assets, and boilerplate that increase network transfer and context tokens without improving an agent’s task state. Janus reframes agent-web optimization as an explicit edge content-negotiation problem: agents opt into machine-oriented representations, while ordinary users continue to receive original pages. The gateway generates, validates, and caches compact semantic variants at the provider or CDN edge, and passes through unsupported content unchanged. Preliminary read-heavy workloads show up to 40.1% task-time reduction, 62.4% context-token reduction, and over 70% fetched-volume reduction for browser-driven agents.
Authors: Zixuan Chen, Zijie Huang
Language: English
Topic: Caching and Content Delivery
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204256418