A Two-Stage Intent Recognition Framework in Multi-Agent Environments

With the rapid development of generative artificial intelligence and the rise of AI agent, accurately identifying user intentions through natural language has become a key challenge in multi-agent systems. Traditional rule-based and classification models for intent recognition face limitations in scalability, semantic understanding and data requirements. Although large pre-trained language models improve accuracy, they often suffer from high computational costs and latency. To address these issues, this paper proposes a two-stage intent recognition framework in multi-agent environments. The first stage employs a lightweight TF-IDF-based approach for efficient initial matching, while the second stage leverages a large language model for fallback recognition when necessary. This design effectively balances speed and accuracy, offering a practical solution for intent recognition in complex, dynamic multi-agent systems.