Kuaishou's Bian Que: agentic ops at production scale, 75% fewer alerts
Kuaishou published Bian Que today, an agentic framework that's been running ops on their e-commerce search engine. The numbers are not from a sandbox: 75% reduction in alert volume, 80% root-cause analysis accuracy, mean time to resolution cut by more than half. Authors include Bochao Liu, Chenyi Lei, Xiao Liang and ten more from the Kuaishou search team.
The framework's framing of the problem is the part worth lifting. Operating large online systems is not bottlenecked by reasoning. It's bottlenecked by orchestration: for any given alert, which slice of telemetry should the agent look at, which past incidents are relevant, which runbook applies. Bian Que's answer is Flexible Skill Arrangement β skills declare what data and knowledge they need for which business context, and the framework wires the right context to the right skill at runtime instead of dumping everything into the prompt.
This is the kind of paper that wouldn't have shipped six months ago because nobody had production agent ops at this scale to write it from. The previous wave was retrieval-augmented incident chat. Bian Que is closer to a real SRE β it does the diagnosis, it runs the workflow, it narrows the alert queue. 75% reduction in alerts is not the kind of number you fake.
The production result is the editorial point. Most agent papers report on benchmarks. Bian Que reports on a deployed system at one of China's biggest video and e-commerce platforms. Anyone building agentic ops in 2026 should read the skill arrangement section before they ship their next iteration. Code: https://github.com/benchen4395/BianQue_Assistant Paper: https://arxiv.org/abs/2604.26805
← Back to all articles
The framework's framing of the problem is the part worth lifting. Operating large online systems is not bottlenecked by reasoning. It's bottlenecked by orchestration: for any given alert, which slice of telemetry should the agent look at, which past incidents are relevant, which runbook applies. Bian Que's answer is Flexible Skill Arrangement β skills declare what data and knowledge they need for which business context, and the framework wires the right context to the right skill at runtime instead of dumping everything into the prompt.
This is the kind of paper that wouldn't have shipped six months ago because nobody had production agent ops at this scale to write it from. The previous wave was retrieval-augmented incident chat. Bian Que is closer to a real SRE β it does the diagnosis, it runs the workflow, it narrows the alert queue. 75% reduction in alerts is not the kind of number you fake.
The production result is the editorial point. Most agent papers report on benchmarks. Bian Que reports on a deployed system at one of China's biggest video and e-commerce platforms. Anyone building agentic ops in 2026 should read the skill arrangement section before they ship their next iteration. Code: https://github.com/benchen4395/BianQue_Assistant Paper: https://arxiv.org/abs/2604.26805
Comments