Zhiyong Kaiwu Raises Nearly RMB 100 Million to Put AI Agents in Factory Roles

Guan told Leiphone the round was the company's third in the past year and that a new round would be completed soon. Zhiyong Kaiwu builds agents as virtual factory employees, including production schedulers, quotation clerks and equipment maintenance engineers. The company says these agents can be onboarded in four steps and used within half an hour. Guan argues that after a decade of misreading the industrial internet, agents are what finally make it start working.

Before founding Zhiyong Kaiwu, Guan spent 19 years at Microsoft, working on IoT, industrial internet and the company's engineering work with OpenAI. He said his most lasting impression came from a factory visit about a decade ago, when he arrived as a technical veteran to explain smart factories and industrial internet platforms. A factory owner asked him, 'What you say sounds good, but what use is it to me?' Guan said he could not answer questions grounded in factory economics, such as how to compare costs when a batch of materials has not arrived and a line is changed or not changed, or how to choose among predictive maintenance options when one shutdown can cost RMB 20 million.

He then proposed a 'loosely coupled' framework, arguing that industrial 4.0 should not follow the consumer internet logic of one big platform solving everything. Specialists should do specialized work, with data connecting their capabilities. The idea hit a bottleneck because no one could explain how to connect experienced workers from different domains. The large model wave changed that. Guan saw a large model as a 'master worker who knows a little about everything,' potentially connecting industrial links that are individually skilled but not interconnected. In 2023, he left Microsoft and founded Zhiyong Kaiwu with former colleagues.

Over three years, the company translated loose coupling into a technical language: multiple agents dynamically collaborate and autonomously decide what to do and how processes should proceed. Guan said a similar logic has been pursued by OpenClaw this year, while Zhiyong Kaiwu has embedded it in industrial sites. The harder part, he said, is not technical validation but entering real factories. He spends 50% of his time on products and 50% on aligning understanding, attending conferences, local workshops and implementation events. He still stays in factories until 11 p.m. or midnight, work he says AI cannot replace. Factory owners ask him what their factories will look like in five years, what role they can play in this wave, and how to implement AI.

By 2026, many companies are still at the knowledge-base stage, piling rules and regulations into an internal Doubao-like repository, Guan said. Many cannot even write clear job descriptions, let alone help AI understand business logic. In that gap, companies can only treat AI as a tool. Zhiyong Kaiwu is translating loose coupling into an executable method. It first uses an industrial semantic engine to help AI understand a company's internal business logic and semantic rules, then lets multiple agents collaborate reliably within real job relationships. The goal is to solve real problems in high-value business domains and help companies move from tool intelligence to organizational intelligence. Guan said this approach is heavier and slower than selling an AI platform, but there is no shortcut for industrial AI; the foundation must reach the semantic layer to avoid another 'big platform' becoming a castle in the air.

Guan's doubts about industrial internet began earlier. Around 2020, while studying for an MBA, he was shaken by a leading factory with a strong digital foundation and high IT investment whose systems failed to help during an emergency. In the preceding years, he had spent time on factory floors. He recalled being unable to answer the owner's cost questions. The mainstream approach was a cloud platform that managed everything. Guan argued that industrial revolutions had advanced through division of labor and skilled specialists, so the fourth industrial revolution should not require one platform to solve all problems. The missing piece was a way to connect specialized processes. The Ministry of Industry and Information Technology liked the loose-coupling idea and asked him to write a white paper, but he again could not answer how much money it would save. Xu Mingqiang, then Microsoft Greater China CTO, suggested he study marginal revenue, which led him to the MBA program.

He now says product development must be moved forward to customer sites. He recently visited electronics and cosmetics factories to help analyze which processes are highly certain and automated, and which consume large amounts of resources, time and labor without being done well. China's equipment factories have strong capacity and run efficiently, but 60% to 70% of their operations may be outside automated lines. In decisions, operations and R&D, specialists know their own work but not whether materials will arrive on time, what factors affect production ramp-up, or whether a product will deform when installed at a customer site. AI can connect their decisions, operations and R&D while specialists continue to do specialized work, which Guan said can bring visible changes in efficiency and cost.

When communicating with business owners and government officials, Guan hears three broad questions: whether they can make judgments about the next five years and how a factory or industry will change; what role they can play if they know the direction, whether to lie flat or strive for more; and how to implement AI. Companies know AI can help but want to know the next step. Guan said the industry lacks products and methodologies that understand both AI and industry and have been validated by very large enterprises.

After leaving Microsoft in 2023, Guan first set up a concept validation center, trying to copy Microsoft's approach of quickly combining existing technology and scenarios. It did not work because people did not understand what the company was doing. Three groups understood and were willing to invest resources, but the company was not Microsoft and some technical capabilities could not meet the scenarios' needs. After ChatGPT emerged, the team tried large models plus bots, and when one bot was not enough, multiple bots, across 20 to 30 scenarios. They concluded that the real worker was not the large model but the agent. Large models are better at reasoning and planning, like the brain's System 2 slow thinking, while the real world needs System 1 immediate reactions, sensor perception, robotic arm control and feedback learning. One agent cannot think, work and perceive at once, so agents must divide labor like people.

Drag-and-drop low-code workflows will not work, Guan said. Business staff cannot draw system prompts and actions, while technical staff do not understand relationships between business nodes or how to handle exceptions. No one can draw one accurate map of these relationships, and there are many engineering problems. Workflow architectures are also fragile. Drag-and-drop is like stacking blocks and can work in digital systems facing deterministic physical processes, but 90% of high-value industrial scenarios are not fixed processes; they require continuous judgment under dynamic constraints, different choices under different environmental variables and feedback based on domain knowledge. If ten business nodes are connected and each pushes million-token context to the next, multi-head attention mechanisms struggle to meet industrial requirements for stability and traceability. Demos may work, but financial data cannot be wrong and industrial sites cannot stop. Drag-and-drop can only do RPA plus some AI judgment, not high-value scenarios. The team judged that path dead.

The industrial semantic engine addresses the problem that AI cannot understand human language while people cannot clearly state their needs. AI appears to process data but does not understand what the data means in a specific business context. For example, the concept of 'material preparation' may be called SKU in a supplier system and material preparation in the company's system; even if they refer to the same physical object, coding rules, field meanings and attribute granularity may differ. Without alignment, data loses key attributes and deforms as it flows. Another problem is local optimization in decision logic. In budget approval, a department requests RMB 2 million in upfront project investment. After the CEO approves, the CFO and sales director see that the top leader has approved and let it pass. No one examines the investment from the company's overall perspective, including whether the input-output ratio meets the bottom line or the highest leverage standard. Without semantic understanding based on global business rules, AI will only mechanically record 'approval passed' and cannot identify risk and reject the application at the first gate. The industrial semantic engine lets AI first understand a company's grammar and logic before working. Guan compared it to AlphaFold, which is first told what a stable structure is and then allowed to think, otherwise it may produce a completely infeasible plan.

After AI understands business logic and aligns semantics, the next question is how multiple agents collaborate. Zhiyong Kaiwu calls its approach a virtual team. Guan said the biggest difference from what the industry calls multi-agent systems is 'innate understanding' versus 'temporary handshaking.' He described some multi-agent architectures in the industry as siloed.