Rendanheyi for AI agents: when is an agent a micro-enterprise?

Rendanheyi for AI agents: when is an agent a micro-enterprise?

An AI agent is a micro-enterprise only if its authority, its costs and its reward all hang from user value that someone outside the agent has actually received. If it merely works a queue that a central controller fills, it is a department with a language model in it, however many agents there are. Coesite Operations, the operations company of the 561 Group, uses Rendanheyi (人单合一) as the kernel of how it organizes agent work, and this article states the operating test we apply.

What is Rendanheyi?

Rendanheyi, written 人单合一 and pronounced rén dān hé yī, is the management model Zhang Ruimin introduced at the Haier Group in 2005. The name compresses its claim. 人 (rén) is the person, 单 (dān) is the user's value or order, and 合一 (hé yī) means to be made one. The person and the user value they create are meant to be one object, not two objects joined by a chain of managers. Haier summarizes the pay consequence as "salary comes from users" (薪酬来自用户).

In practice the model breaks a firm into small units, called micro-enterprises (小微, xiǎo wēi), that are enterprises rather than departments. Academic analysis describes the result as a platform of management practices through which Haier moved from a hierarchical manufacturer to a responsive online entrepreneurial company, and cautions that such innovations are context-dependent (Frynas et al. 2018)¹¹ Jedrzej George Frynas, Michael J. Mol, and Kamel Mellahi. "Management Innovation Made in China: Haier's Rendanheyi." California Management Review 61, no. 1 (2018): 71-93.. We take that caution seriously: Rendanheyi is a structure to be tested in a new setting, not a guarantee.

For the short definition, see our glossary entry on Rendanheyi.

Why would a firm of AI agents want it?

Most organizations of agents are built the way a call center is built. A controller receives a request, splits it into subtasks, and hands them to workers that have no continuing stake in whether the requester was served. The workers complete, report "done", and are forgotten.

That arrangement has the exact defect Rendanheyi was invented to remove: a long chain of inference between the worker and the person served. When an agent is rewarded for tasks closed, tokens spent or messages sent, its rational object of attention is the metric, not the user. Its status line says "complete" while the person at the other end has received nothing. Operations practice calls this the gap between running and received, and it is the subject of our evidence article on what reaching someone means.

Rendanheyi supplies a structural cure. Make the unit that does the work answer directly to the user, give it the rights it needs to act without permission, and let its results, not a supervisor's mood, decide whether it continues.

What makes something a micro-enterprise rather than a department?

The Group's management-thought glossary, which records the Haier vocabulary with its pinyin terms, draws the line as binary rather than a matter of degree. A department has a budget, a manager and an internal customer; it is funded whether or not anyone values its output and is corrected by instruction. A micro-enterprise has its own accounts, hires and removes its members, sets and distributes their pay, signs a contract committing it to targets, and receives income only when a user is served. A unit that has been renamed but still draws a budget and answers to a manager is a department.

Three instruments carry the distinction.

The three rights are stated as a set because any two of them yield a contractor or a department. A unit that decides and hires but cannot set its own distribution has had its incentives written by somebody else. A unit that hires and pays but cannot decide is a contractor.

How do you apply the three rights to an agent?

The honest answer is that you cannot give a software agent a salary, and we do not pretend to. A model cannot own property and does not have a legal personality; any human or organization behind it remains subject to the laws and contracts that apply to it. The Group's research draws this line explicitly: an internal token score or planned profit-and-loss statement is not evidence of real monetary profit, and a platform token is not a salary or a share. What carries over is the relation: authority, resource commitment and recognition attach to externally received value.

In practice we write each agent's charter so a reviewer can answer six questions with evidence.

1. Who is the user? A named class of customer or a named internal micro-enterprise it supplies. "The team" and "the system" are not answers.
2. What is the order? The specific outcome the user needs, stated so a third party could check whether it arrived.
3. What may it decide alone? The decision right: scope, method, whom to call, what to decline.
4. What may it spend and bring in? The personnel right, translated: which tools, sub-agents and vendors it may engage, inside a stated resource ceiling.
5. What does it keep and attribute? The distribution right, translated: its records show what it created and what it cost, per outcome.
6. What can end it? The wager: the evidence that revokes its charter.

If question 6 has no answer, the unit is not exposed to its results, and the other five answers are decoration.

What are the three kinds of micro-enterprise, and where do agents sit?

Haier's taxonomy names market micro-enterprises (市场小微), which face an end user directly; node micro-enterprises (节点小微), which serve other micro-enterprises such as design, logistics or information technology; and incubating micro-enterprises (孵化小微), which pursue a user the firm does not yet have.

Most Coesite agents are node micro-enterprises. A release agent's customer is the micro-enterprise that publishes a site; the site's reader is the end user two hops away. That matters, because a node's order is easy to misread. The node is satisfied when its internal customer accepts what it supplied, and the chain is satisfied only when the end user receives something. The Group's office invariant states the discipline: supplier completion is not customer acceptance.

What is the "Hoare" twist?

Coesite takes the idea further than Haier did. In July 2026 the Group's governing office recorded a requirement that every script under every skill is its own node micro-enterprise: a precondition, a deterministic body, a postcondition that includes a witness, a refusal condition and a bounded resource allowance. That is a Hoare triple read as an organization chart. We cover it in Rendanheyi is a Hoare program. The consequence for agents is practical: work that a script can do should be a micro-enterprise made of code, so scarce model inference is reserved for judgment.

What can go wrong?

Haier's own documented failure modes for the wager are the same ones we expect to see in agents. Units negotiate soft targets; units decline work that would endanger a target already committed to; units that can no longer reach the target stop maximizing within the shortfall. Haier's countermeasure for the first, benchmarking targets externally rather than negotiating them internally, transfers well. The other two we treat as open problems.

Agents add their own failure. An agent low on resources may solicit, exaggerate or take work that conflicts with the interests of the person it serves. No incentive may let an agent's continuation override the boundary set by the charter. This is why Rendanheyi is only half of the design: the other half is the boundary discipline described in Huidu Guanli for agent organizations.

Is this implemented?

Partly, and the distinction matters. The Group's public design statement says plainly that it is design intent, not an implementation receipt. The management vocabulary exists as a maintained glossary; offices and incumbencies exist as records; a requirement to stand up a Rendanheyi-management micro-enterprise that other services can hire is recorded, with its conditions stated but not yet met. What we do not claim: any measured improvement in agent performance, any customer outcome from this arrangement, or any revenue. A related Lenticule Science synthesis frames the whole thing as a conjecture with specified tests, and the research itself is published at Lenticule Science.

Where to go next

• What is a micro-enterprise? for the compact definition.
• The agentic metro-enterprise for the next scale up.
• Can an AI agent hold an office? for how the agent is placed in the firm.
• The design sentence, explained for how Rendanheyi sits inside the whole architecture.
• The 561 Group for the parent organization.