AgentPPT: an LLM agentic approach for autonomous PowerPoint operation

Abstract

Recent advancements in Large Language Models (LLMs) have demonstrated their capabilities across complex domains such as logic, mathematics, and programming, enabling the development of autonomous agents for a wide range of applications. This project explores the novel application of LLM-based agents for operating application software, specifically Microsoft PowerPoint. Building on the PPTC framework, which enables LLMs to create and edit presentations through natural language commands, this work identifies and resolves a critical evaluation error in the original system and proposes architectural enhancements for improved performance. Furthermore, we introduce AgentPPT, a new LLM-driven agent that separates planning and execution phases to increase task efficiency and reduce redundancy. A user interface is also developed to facilitate interactive presentation editing using either PPTC or AgentPPT. The proposed contributions aim to bridge the gap between content generation and autonomous operation of presentation software, offering a more robust and efficient solution for real-world use.