Feihe built 436 AI agents across 10+ dairy plants. Most came from a factory-floor contest.

Feihe (中国飞鹤), China's largest infant formula maker, runs more than ten modern dairy plants in the country. On August 27, the company and ByteDance's enterprise cloud unit, Volcano Engine (火山引擎), published a co-branded case study claiming 436 AI agents now operate across those plants, with average equipment fault recovery time cut from 104 hours to 47 hours. The most concrete part of the report is not the scale figure. It is how those agents were chosen: in a sponsored factory-floor contest, twelve finalist projects were proposed mostly by line workers and maintenance staff, not by the IT department.

The deployment sits on four pieces of ByteDance infrastructure. TRAE is a low-code environment for building agents without writing code. HiAgent is the unified agent platform where the agents actually run. Arkclaw orchestrates actions across Feihe's existing enterprise systems. Doubao, ByteDance's large language model family, provides the language and reasoning layer. Together, the stack is the load-bearing piece: every factory in the program pulls from the same agent registry, the same low-code environment, and the same orchestration layer, which is how a single maintenance procedure can be packaged once and shipped to every plant. (Leiphone)

The first deployment to produce hard numbers is AI TPM, an "intelligent equipment management operating system" now running across nine Feihe plants and serving more than 1,000 technical engineers. The case study reports that average fault recovery dropped from 104 hours to 47 hours, and the repeat-fault rate fell from 34 percent to 8 percent. The platform logs roughly 23 billion tokens of LLM traffic and around 240,000 active users over a three-month window. (Leiphone)

A second deployment, called 精工膜创 (precision membrane filtration), sits inside a single factory's milk-protein separation process. More precise cleaning control extended membrane life, saving roughly 3 million yuan (about $420,000) in equipment replacement, cutting annual procurement of filter consumables by more than 500,000 yuan (about $70,000), and reducing annual water use by about 144,000 tons. A third project, 鹤勤智联, rebuilt a human-resources and workforce workflow, compressed fifteen process steps into five, and reported a more-than-threefold efficiency gain alongside more than 2 million yuan (about $280,000) in cumulative development and operations cost. (Leiphone)

The 2026 AI Innovation Contest (Factory Season) is the part of the program that generalizes. Feihe says the contest ran across more than ten factories, drew more than fifty teams, and produced twelve finalist projects. Most scenarios came from frontline employees, a maintenance technician flagging a recurring pump fault, a quality-control worker wanting faster batch-document search, a line operator trying to reduce manual cleaning cycles. The pattern matters because the contest is the intake funnel: it produces a steady stream of agent candidates that the unified platform then hosts. Without that funnel, the 436-agent figure has no engine behind it. (Leiphone)

Feihe's underlying factory base is real. The company operates more than ten modern smart factories with a digital production equipment share above 90 percent, and a 2024 Nandu report placed its production capacity at eleven facilities with more than 363,000 tons per year of design output. Feihe is listed in Hong Kong as 6186.HK, and basic listing data is available through Xueqiu, Sina Finance, Yahoo Finance, Reuters, and Bloomberg. The 2024 production-capacity figure is from Nandu.

Every quantitative outcome in this article comes from one Leiphone article sourced to Feihe executives: the 104-to-47-hour recovery cut, the 34-to-8 percent repeat-fault drop, the 3 million yuan membrane-life saving, the more-than-500,000-yuan annual procurement saving, the 144,000 tons of water, the more-than-2-million-yuan workforce-system saving, and the 23 billion tokens. Named sources include VP of Production Sun Jianguo (孙建国), Digital Growth Center head Zhi Qiang (只强), AI lead Nan Ding (南丁), and AI TPM developer Zu Xiuxiu (祖秀秀). Feihe and Volcano Engine both have a stake in this being told as a success. The number worth testing is not the 436-agent tally. It is whether the pattern, a factory-floor contest feeding a unified agent platform on a single cloud, produces the same results at a Yili or a Mengniu plant, on a different cloud, or under a different contest sponsor.

USD figures here use an approximate 7.2 RMB per USD rate and are labeled as such; the original RMB amounts remain the auditable source figures.