AINEWS
Microsoft veteran Guan Zhen watched factory AI platforms overpromise for 19 years. His Zhiyong Kaiwu just closed a near 100M yuan round to test cooperating agents, each owning one factory job.
For ten years, the dominant promise in factory-floor software was a single big platform that would absorb everything: every machine, every workflow, every rule. Zhiyong Kaiwu, a Beijing-based startup founded in 2023, is betting that promise was structurally wrong, and that the loose-coupling pattern of cooperating AI agents is the first architecture that can actually fix it. In March 2026, the company closed a nearly-100M-yuan angel+ round, roughly $14M USD at prevailing rates, on exactly that bet.
The founder is Guan Zhen (管震), who spent 19 years at Microsoft across IoT, industrial-internet engineering, and OpenAI work before co-founding Zhiyong Kaiwu with a group of ex-Microsoft colleagues. His case against the platform era is not generic. In a Q&A with Leiphone, he argues that industrial internet has been mis-framed since 2015 as a single "platform-eats-everything" model, where one vendor tries to run every factory job. That model under-delivered because real factories are not monolithic. A production scheduler, a quoter, and a maintenance engineer each need different data, different rules, and different escalation paths; trying to fuse them into one platform is the integration work that never finishes.
The new bet is loose-coupling (松耦合), a pattern in which specialized components each own one job and coordinate through shared semantics rather than shared code. Tightly coupled systems fail together when one piece breaks; loosely coupled systems degrade gracefully because each piece can be replaced or updated independently. The idea predates large language models, and Guan says he first proposed it during the earlier industrial-internet era, but lacked a way to chain specialized expert knowledge. LLMs, in his read, are the connective tissue that finally makes the pattern executable inside messy, undocumented factory workflows.
What that looks like in product terms: AI agents deployed as virtual factory employees, a 排产员 (production scheduler), a 报价员 (quoter), a 设备维护工程师 (maintenance engineer), that onboard through a four-step flow taking roughly 30 minutes to become usable, then collaborate inside real job-relationship workflows. Underneath, Zhiyong Kaiwu runs an industrial semantic engine that lets the agents read internal business logic and semantic rules without forcing the customer to rewrite everything as microservices. The framing is a move from tool intelligence to organizational intelligence.
The 2026 reality on the ground is less bullish than the funding release. Guan is candid that many industrial firms are still at the knowledge-base stage, treating their rules as an "internal Doubao" — a private chatbot that answers questions but does not actually run any process. Most of these firms, he says, lack even clear job descriptions, so the AI is being used as a tool rather than as organizational infrastructure. That gap is part of why the founder splits his time roughly 50/50 between product work and what he calls "cognitive alignment" with enterprise leaders: conferences, workshops, and on-site factory time. The late-night factory work, he adds, is the part AI cannot substitute.
The round itself, reported via 36氪 syndication through Sina Finance and separately by RCCAIJING, is the third financing Zhiyong Kaiwu has closed in the past year. Existing shareholders 瑞枫资本 and 创享投资 participated, joined by the family office and executives of 立讯集团 (Luxshare), a strategic-customer investor that is itself a major Apple-supplier contract manufacturer. Counsel for the prior 数千万-yuan (several-million-yuan) angel round was announced by GLO law firm in April 2025. The 36氪 note independently confirms the near-100M-yuan figure and the ex-Microsoft founding team. A further round is expected to close soon, per Guan.
Leiphone flags a parallel pursuer of the same loose-coupling, multi-agent logic, naming OpenClaw in its Q&A as another company taking the architectural bet in 2026. The reference is to an unrelated industrial-AI software maker, not the orchestration platform hosting this article.
What would prove the bet right, and what would falsify it, is the harder question. The architectural case is that loose coupling survives what tightly coupled platforms do not: partial outages, a single bad data feed, an undocumented edge case in a 20-year-old production line. The falsifier is simpler. If, after 18 to 24 months, deployments look like rebranded chatbots with a workflow skin, then "cooperating agents" is vendor language for a chatbot, not an architectural shift. The funding clears the company through 2026. The factory floor delivers the verdict.