Not All AI Agents Are Equal: Characterizing Resource and Performance Dynamics

Concurrent agent execution creates unpredictable latency due to the inter-mixing of remote LLM API calls and local tool container demands. Characterization of RAG, web search, and coding tasks reveals that resource dynamics vary significantly by task, even when using identical tools. This variance complicates serving optimization in multi-request environments.