LLD-Agent: Enhance the runtime efficiency of agents with decoupling the execution phase from the creation phase

Abstract

Large Language Models(LLMs) based AI Agents have demonstrated strong capabilities across a wide range of domains. However, conventional agent frameworks rely heavily on LLMs throughout both the creation and execution phases, leading to low efficiency and nondeterministic outcomes. To address these limitations, this paper proposes the LLD-Agent platform. By designing an App Script and introducing the Memory Bus and Glue Code mechanisms, LLD-Agent decouples agent construction from execution, enabling the execution phase to avoid repeated LLM-based task planning and tool selection. Specifically, the App Script predefines and persists the agent workflow; the Memory Bus supports information sharing during execution; and Glue Code enables advanced customization of the execution pipeline. Experimental results show that LLD-Agent significantly outperforms other agent platforms and function-calling LLM baselines in runtime efficiency, validating its effectiveness in improving agent execution performance. This study provides a new approach and practical evidence for building efficient and controllable AI Agents platforms.

Ähnliche Arbeiten