Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations
Researchers have introduced Bian Que, a novel agentic framework designed to streamline the operation and maintenance (O&M) of large-scale online engine systems, such as search and recommendation engines. Addressing the bottleneck of orchestration in LLM-based agents, Bian Que focuses on the precise selection of relevant data and operational knowledge rather than just reasoning capabilities. The framework features three key contributions: a unified operational paradigm abstracting O&M actions into release interception, proactive inspection, and root cause analysis; a flexible Skill Arrangement mechanism where skills defining requisite data and knowledge are automatically generated or optimized via natural language; and a self-evolving mechanism that distills event memory into knowledge. Deployed on KuaiShou’s e-commerce search engine, Bian Que demonstrated significant performance improvements, including a 75% reduction in alert volume, 80% accuracy in root-cause analysis, and over 50% reduction in mean time to resolution. The system also achieved a 99.0% pass rate in offline evaluations. This development highlights the practical application of AI agents in complex industrial IT operations, offering a scalable solution to reduce human effort in monitoring and troubleshooting high-frequency release environments.
Wire timeline
Bian Que: An Agentic Framework with Flexible Skill Arrangement for Online System Operations
Researchers have introduced Bian Que, a novel agentic framework designed to streamline the operation and maintenance (O&M) of large-scale online engine systems, such as search and recommendation engines. Addressing the bottleneck of orchestration in LLM-based agents, Bian Que focuses on the precise selection of relevant data and operational knowledge rather than just reasoning capabilities. The framework features three key contributions: a unified operational paradigm abstracting O&M actions into release interception, proactive inspection, and root cause analysis; a flexible Skill Arrangement mechanism where skills defining requisite data and knowledge are automatically generated or optimized via natural language; and a self-evolving mechanism that distills event memory into knowledge. Deployed on KuaiShou’s e-commerce search engine, Bian Que demonstrated significant performance improvements, including a 75% reduction in alert volume, 80% accuracy in root-cause analysis, and over 50% reduction in mean time to resolution. The system also achieved a 99.0% pass rate in offline evaluations. This development highlights the practical application of AI agents in complex industrial IT operations, offering a scalable solution to reduce human effort in monitoring and troubleshooting high-frequency release environments.
cs.AI updates on arXiv.org