Intelligent Agent Technology in Commercial Advertising Platforms: Architecture and Applications

This article explores the application of intelligent agent technology in commercial advertising platforms, focusing on the architecture and implementation of multi-agent systems powered by large language models (LLMs). The authors from Baidu's Commercial Advertising Platform R&D team discuss how AI-native advertising platforms are undergoing fundamental transformations, with intelligent agents serving as the primary vehicle for delivering commercial value to customers.

The article outlines four core capabilities that intelligent agents should possess: understanding natural language queries, proactive planning, strong execution capabilities, and human-like responses. It then delves into the technical challenges faced when applying LLM agent technology to advertising platforms, including accurate query parsing, autonomous planning limitations, business system integration difficulties, and natural language response generation.

The authors present Baidu's multi-agent architecture based on Wenxin (ERNIE) 4.0, which addresses these challenges through three key innovations: a 'large-small model collaborative' architecture that leverages long-term memory to solve understanding and response latency issues; domain SOP-based multi-agent collaboration to handle autonomous planning and execution weaknesses; and a long-term memory plus self-learning strategy to optimize the data flywheel and drive incremental customer expression.

The architecture consists of five layers: application layer (assembling vertical domain agents), agent layer (framework infrastructure and multi-agent collaboration), model layer (large and small models with supporting tools), memory layer (vector data and long-term memory stored in BaikalDB), and data toolset (evaluation, testing, and annotation tools).

Two practical applications are detailed: the GBI intelligent agent for natural language-based business intelligence analysis, and JarvisBot for intelligent operations and fault diagnosis. The article concludes with the benefits achieved through these implementations, including increased consumption, improved accuracy and speed of LUI interactions, and significant efficiency gains in both business and engineering operations.

The authors also share their technical insights and reflections on LLM hallucination problems, the need for comprehensive vector databases, and the future potential of multimodal and cross-modal models in advertising marketing.