Dynamic Service Performance Considerations for AI Agent Selection (draft-zhu-agent-dynamic-service-selection-00)

Abstract

In the Internet of Agents, registration and capability discovery can identify candidate agents that meet basic functional constraints. However, the response performance of candidate services may vary with the underlying model, concurrent requests, and deployment environment. Static capability declarations or generic computing- load indicators alone may not adequately represent actual agent service performance. This document discusses the applicability of dynamic information, including Time to First Token (TTFT), output token generation rate, request success rate, concurrency, and availability, to candidate agent selection. It focuses on the observed entity, data sources, measurement windows, information freshness, and prerequisites for comparing candidates, and outlines an illustrative selection process.