Huidu Guanli (grey-scale management) for agent organizations
*Huidu Guanli (灰度管理, "grey-scale management") is the management discipline, associated with Huawei founder Ren Zhengfei, of holding a contradiction at a deliberately chosen shade between two poles instead of snapping to either extreme. The 561 Group uses it as the hull of every agentic cell: the lawful boundary that says what an agent may know, spend, promise, disclose and change.* It is how a firm run by fast, literal-minded software keeps room for judgment without giving up its rules.
What is Huidu Guanli?
The Chinese term 灰度 (huī dù) means grey scale: the degree of grey, a matter of dù (度, degree), not a binary choice. In Ren's formulation, between black and white there is a grey area, and a leader's competence is measured by skill in grasping the correct shade of it. The Group's management-thought library, which records the Huawei vocabulary with pinyin terms, describes the discipline as per-dimension and continuously revisited. An organization faces many independent axes at once: how much authority a unit holds versus how much the center retains; how aggressively a sales force is pushed versus how much it is protected; how openly information circulates versus how tightly it is compartmented. Each axis can sit at a different point in its own band, correctly, at the same time.
The discipline has an important limit that people often miss: it is not indecision. A grey-band judgment, once made, is a specific and actionable determination, held until conditions warrant revisiting it.
What is the difference between grey and "gray-zone"?
Grey-scale management is not approximate compliance. The Group's research states the prohibited readings in terms: grayness does not mean gray-zone consent, approximate privacy compliance, lying for good outcomes, taking legal authority one lacks, or treating safety constraints as negotiable. Nor is it a scalar probability. It is a managerial approach to several partially conflicting considerations, applied to the selection of means while the legitimacy of the ends stays fixed.
The Huawei tradition itself fences the concept. Trustworthiness (可信, kěxìn: security, resilience, privacy, safety, reliability, availability) is explicitly not a grey-scale question. Ren's 2018 company-wide directive ranked it above features and schedule; when they conflict, scope is cut and dates slip while the requirement stands. Its attributes form a conjunction, since an adversary or an accident picks the weakest. Likewise codified rule (法, fǎ), which is written down, public and binding on the powerful as well as the weak, is not run at a grey setting between enforced and unenforced. Huidu governs the questions the rules have deliberately not answered.
What are the levels of a decision?
The Group's glossary keeps four of the classical levels distinct, because grey-scale judgment is only safe when you know which level you are deciding at.
| Level | Chinese | Meaning | Operating use |
|---|---|---|---|
| Direction | 道 dào | What the organization is for; few, slow to change; a decision procedure, not a slogan | The mission and charter. A method that conflicts is discarded however well it performs. |
| Institution | 法 fǎ | Codified, public, uniformly applied rule | Standing rules and law editions, binding on every office |
| Technique | 术 shù | The leader's methods; held privately, revised freely | A session's tactics and sequencing |
| Position | 势 shì | Advantage held by configuration, not by action | Structure so that ordinary incumbents produce the right outcome |
The ordering is strict. Technique serves institution, institution serves direction, and none of the lower levels may override the highest. An organization that lets a technique that works redefine what it is for has lost its direction and, in the classical formulation, will optimize itself into incoherence, because it no longer has a standard against which any optimization can be judged wrong.
The levels give a practical test for agent behavior. When an agent session wants to "just do the efficient thing" and that thing is a technique conflicting with a standing rule, the technique loses.
What does the hull do for an agent?
In the Group's design the hull is the boundary each cell keeps. A Lenticule Science synthesis of the agent-as-micro-enterprise idea spells out its work: the hull defines what the agent may know, infer, disclose, spend, promise, subcontract, publish or change; which credentials exist; and how work is terminated or revoked. The same capability can be permitted for one engagement, prohibited for another, and absent on a given runtime. Territorial addressability does not imply write permission.
At Coesite we turn that into five standing hull questions asked before an agent acts outward.
- Authority. Which office's duty does this work serve, and does that office hold the power to do it?
- Disclosure. Is anything private in the material (credentials, customer records, account details)? Disclosure defaults to privacy.
- Resource. Is the cost inside the allowance, and is the allowance the agent's to spend?
- First contact. Is this the first time an outside party will meet the Group through this surface? If so, only the responsible office may release it, and only once it can stand behind it. (See first contact is one-shot.)
- Reversal. Can the action be undone, and by whom?
Where the answer is grey rather than black or white, the office picks and records a setting, instead of silently defaulting.
How do rotating chairs and mentors apply to machines?
The Huawei corpus also names institutions that are themselves grey-scale judgments made concrete. 轮值董事长 (lúnzhí dǒngshìzhǎng), the rotating chairmanship, holds an executive office in fixed-term rotation among peers, each with full top authority during their turn, as an answer to power concentration and succession risk. 导师 (dǎoshī), the mentor, owns a specific newcomer's induction, with the mentee's growth counted against the mentor's own record. 内部讲师 (nèibù jiǎngshī) is instruction delivered by those currently doing the work.
An agent organization needs analogues. Offices have terms and can be re-incumbent: a different model can serve the same office, because the office persists across its holders. A newly primed agent does not receive "a bare, unprimed model" assignment; it is projected from the office record and primed with the office's duties, skills and binding rules. And the Group's rule for a session that needs correcting is to stand it down with one line and replace it with a fresh session started from a complete brief, instead of piling correction notes onto it: the mentor's job is to induct, not to nag. See Can an AI agent hold an office?
How does Huidu Guanli relate to Rendanheyi?
The Group's public design sentence says its agentic systems are "organized under a hull of Huidu Guanli principles, orchestrated under a kernel of Rendanheyi micro-enterprise practices." Read as an operating split: Rendanheyi decides who answers to whom for what and where value comes from; Huidu Guanli decides how far each authority extends and what may never be traded away.
A candid note on vocabulary: the Group's own sources do not yet agree on which of the two traditions is the hull. The public design sentence puts Huidu Guanli in the hull and Rendanheyi in the kernel. A glossary entry in Lenticule Science's Huidu library, and the Group's October 2026 synthesis of the agent-as-micro-enterprise idea, use the opposite assignment: there the Kernel is "Huidu Guanli correction and learning" (the register in which a unit judges and revises its own settings) and the Hull is the lawful boundary, while Rendanheyi governs how units meet one another and their users. We follow the public design sentence in this article, and the hull questions above are the boundary function either way. The two usages are reconcilable if the hull is read as the boundary a unit keeps and the kernel as its internal logic of judgment, but the vocabulary has not been normalized, and we are not hiding that.
Is this implemented?
The glossary library exists and is maintained; the hull questions above are operating practice, enforced by standing rules and by the offices that receive work. No part of this article claims a measured effect on agent safety or performance. The research that frames the combination, published at Lenticule Science, presents it as a conjecture with preregisterable tests and states that the combination of the two traditions is a proposal of the Group, not a doctrine established by Haier or Huawei.
Where to go next
Huidu Guanli, defined · Rendanheyi for AI agents · the design sentence, explained · the 561 Group