RESOURCE OPTIMIZATION AND COGNITIVE RESILIENCE IN AUTONOMOUS AGENTS VIA THE ACTION-RAG PARADIGM

Abstract

The scalability of autonomous artificial intelligence models is currently restricted by API volatility and the saturation of the Transformer attention window, in which the exhaustive injection of tools causes the "Lost in the Middle" phenomenon. This paper introduces an Action-RAG paradigm, which is based on Retrieval-Augmented Generation(RAG) methodologies and differs through its dynamic capability management, using the MCP protocol to transition from passive injection to an iterative activation state. The purpose of the proposed method is also to optimize the "cost per token" during the search for action tools. This method is accompanied by validation through tests on the ToolBench and Gorilla APIBench corpora (N=14.297). The architecture demonstrated to outperform methodologies such as Passive-RAG by up to 10.6 percentage points in Hit Rate, with an efficiency 31% higher than SemanticKernel.