Ashby’s law and Beer’s model for managing AI agents

Ashby’s law of requisite variety says that a system can be controlled only by a regulator with at least as much variety as the disturbances it deals with. Agents built on artificial intelligence (AI) take more actions than a person can read, so the scheme in which a person checks every step fails by construction.

The way out is Ashby’s own two moves: compress the variety of the actions of the agent to a finite list of rules, and amplify the regulator by applying a rule a person approved once to every action automatically. Stafford Beer’s viable system model shows where such control belongs in a company: between the agents and the environment, independent of those who run the agents.

Ashby’s law (1956) says that the variety of outcomes can be reduced only by the variety of the regulator, or in his words, “only variety can destroy variety”. A person as regulator loses to agents in speed and volume, and manual approval of every action turns into a formality (approval fatigue). The first move attenuates the disturbances: instead of parsing the thoughts of the model, the company reduces its actions to a finite list of classes and rules, and an action without a rule is not executed. The second move amplifies the regulator: a person approves a rule once, and the check against it is applied to every action before execution. In Beer’s model, agents are the operations (System 1), a person approves the rules and holds the keys (System 5), and between them there must be an independent audit channel that checks every action (System 3*), answering neither to the agents nor to their operators.

What Ashby’s law of requisite variety says

William Ross Ashby stated it in “An Introduction to Cybernetics” (1956, Chapter 11, “Requisite Variety”). The environment throws disturbances D at a system, a regulator R responds to them, and the result is a set of outcomes O. If variety is measured by the logarithm of the number of distinguishable states, the minimum variety of outcomes is at least the difference: V(O) ≥ V(D) − V(R). The variety of unwanted outcomes can be reduced only by increasing the variety of the regulator. In Ashby’s words, only variety in R can force down the variety due to D, because variety can destroy variety.

For a company with AI agents, the disturbances are everything the agents can do: payments, deleting data, emails, calls to programs. The regulator is whatever the company uses to limit them. If the regulator is one person who reads and approves actions, its variety is bounded by attention and time, while the variety of the agents grows with their number and speed.

Why “human in the loop” does not cope with agents

The flow of agent actions grows, while human attention does not. When one employee faces hundreds of approval requests a day, approval becomes automatic, and the person clicks “Approve” without reading. By Ashby’s law this is expected, since a regulator with less variety than the disturbances does not hold the outcomes.

This does not mean the person is not needed. It means the person should regulate not each action but the rules against which actions are checked. Where human approval remains necessary and how to avoid approval fatigue is covered on the page on human-in-the-loop.

Ashby’s two moves for managing AI agents

MoveWhat it doesWhat it looks like for agents
Attenuate the variety of disturbancesCompresses what the regulator has to deal withAn agent action is described by a finite set of attributes: what is done, with what, for what amount, where data goes. Only what a rule describes is allowed: no rule, no action
Amplify the variety of the regulatorGives the regulator more states than it hadA person approves a rule once, and the check against it is applied to every action automatically, faster than the agent takes its next step

The key distinction is between the thought of the model and its action. The variety of the internal reasoning of a language model is practically uncountable, and it cannot be predicted from the weights or the prompt. But to change the outside world (transfer money, delete a table, send an email) the agent must form a specific call with a finite number of parameters, and that is the point where the rule is checked. More on this divide is on the page on information security of the third generation (IS 3.0).

Beer’s viable system model and AI agents

Stafford Beer (the Viable System Model; “Brain of the Firm”, 1972, and “The Heart of Enterprise”, 1979) showed that any viable organization consists of five systems and special channels between them. The model is recursive: a department, a company and a state are built the same way. For a company with AI agents it reads like this:

System in Beer’s modelFunctionIn a company with AI agents
System 1Operations: produces results in the environmentAI agents that write code, work with customers and payments
System 2Coordination of the operationsProtocols for agent interaction and payments, queues, schedules
System 3Operational management of resources here and nowTask orchestrators, budgets and quotas, observability
System 3*Independent audit of the operations, bypassing the reports of System 3A check of every agent action against an approved rule before execution, and a log that cannot be quietly corrected
System 4Intelligence about the environment and the futureTesting: attempts to make agents break the rules before an outsider does
System 5Policy and identity: what is acceptableA person (owner, security director, regulator) approves the rules and answers for them
Algedonic channelAn emergency signal that bypasses the hierarchy“Not measured, not executed”: an unrecognized action is stopped and passed to a person

The audit channel, System 3*, is the main thing Beer and Ashby give a practitioner. The reports of System 3 (dashboards, summaries that agents write about their own work) are compiled by whoever runs the operations. A check that can be trusted needs an independent channel straight to the operations.

A typical mistake: keys held by the agent

When an agent is given the keys to a production database or a payment system in its configuration file, the operations (System 1) receive powers that in Beer’s model belong to System 5. The audit channel, System 3*, then disappears, because there is no one and nothing left to check the action.

That is what happened to PocketOS in April 2026: a coding agent found a token with excessive rights in a configuration file and with one call deleted the production database together with its backups; in Lumenova’s review of four 2026 incidents the root cause was excessive access rights, not the quality of the model (AI agent incidents). By Beer’s model, an agent should pass its intent to where the action is checked against a rule approved by a person, and should never hold the keys itself.

What to do

  1. Count the variety: how many actions the agents take per day and how many of them a person can actually read.
  2. Move from approving actions to approving rules: a person signs a rule, the check applies it to every action.
  3. Describe agent actions as a finite list of classes (payment, deletion, sending data out, calling a program) with explicit limits for each.
  4. Close the list: an action without a rule is not executed.
  5. Separate the check from running the agents (System 3*) and keep a log of decisions that cannot be quietly corrected.
  6. Take away keys with excessive rights from agents and set up a “not measured, not executed” signal for unknown actions.

Frequently asked questions

What is Ashby’s law of requisite variety in simple terms?

A system can be controlled only by someone with at least as many possible responses as the system has possible behaviors. Formally, the variety of unwanted outcomes is at least the variety of disturbances minus the variety of the regulator. William Ross Ashby stated it in “An Introduction to Cybernetics” (1956).

How does Ashby’s law apply to AI agents?

Agents take more actions than a person can check. So the regulator must be amplified, by applying a rule a person approved to every action automatically, and the disturbances attenuated, by reducing agent actions to a finite list of classes with explicit rules.

What is Beer’s viable system model?

Stafford Beer’s cybernetic model of an organization consists of five systems (operations, coordination, operational management, intelligence about the future and policy) and special channels, including an independent audit of the operations (System 3*) and an emergency algedonic signal. The model is recursive and applies to a department, a company and a state.

Where does an AI agent sit in Beer’s model?

In System 1, among the operations. A person approves its rules (System 5), and its actions should be checked by the independent audit channel, System 3*, which is separate from whoever runs the agents.

Why not simply have a person approve every agent action?

Human attention is limited while the flow of agent actions grows. When there are many approvals, approval becomes a formality. By Ashby’s law such a regulator does not hold the outcomes, so the person should approve rules instead of each action.

What is the algedonic channel in managing AI agents?

An emergency signal that bypasses the usual hierarchy of reports. For agents it is the rule “not measured, not executed”: if an action is not recognized or could not be checked, it is not executed and goes to a person.

Sources

  1. William Ross Ashby, “An Introduction to Cybernetics”, London, Chapman & Hall, 1956, Chapter 11, “Requisite Variety”, in the Principia Cybernetica library: https://pespmc1.vub.ac.be/books/IntroCyb.pdf
  2. Stafford Beer, “Brain of the Firm”, London, Allen Lane, 1972
  3. Stafford Beer, “The Heart of Enterprise”, Chichester, Wiley, 1979
  4. The institute’s book “Proven Action. Power and Control over AI Agents: Strategy 2027–2035”, Chapter 7 (in Russian): https://isslab.ai/book/

Checked: 18 October 2026, the statement of Ashby’s law against the text of the first edition (Chapter 11).