The vocabulary of agents, harnesses, habitats, roles, and handoffs overlaps with several older vocabularies from control theory, cybernetics, distributed systems, and safety engineering.

The overlap is useful because these fields have spent a long time asking structurally similar questions. What can an actor observe? What actions are available? Which states can the system reach? What keeps a trajectory inside acceptable bounds? When should one controller yield to another? How does a system detect error, recover, and change its own policy?

But the correspondence is not exact. Control theory describes what a system can cause. Institutional governance also asks what an actor is entitled to cause, who must answer for it, and when an action becomes legitimate.

That difference matters.

The Basic Control-System Frame

A minimal feedback-control loop contains a few familiar objects:

                    observations
             ┌────────────────────────┐
             │                        │
             ▼                        │
Goal ──▶ Controller ──▶ Actuator ──▶ Plant
             ▲                        │
             │                        │
             └──── state estimate ◀── Sensor

The plant is the system being acted upon. It may be a codebase, an application, an operating process, a market, or part of an organization.

The controller selects an action based on a goal and its estimate of the current state. The actuators are the mechanisms through which that action changes the plant. The sensors expose what happened. The feedback loop uses those observations to decide what to do next.

An agentic system adds institutional structure around this loop:

Role and task
      │
      ▼
Agent through Harness ──▶ Tools ──▶ Organizational system
      ▲                                      │
      └──────── evidence and feedback ◀──────┘

      all operating inside a Habitat
      and crossing boundaries through Handoffs

The agent and harness together resemble a controller. Tools resemble actuators and sensors. The organizational system resembles the plant. The habitat constrains the admissible loop. Handoff policies govern transitions between controllers, roles, and modes of operation.

This is the rough correspondence. The useful details are in where it stops being rough.

Conceptual Crosswalk

Agentic organization term Control theory or cybernetics analogue What the comparison reveals Where it is incomplete
Organization Plant plus a hierarchy of controllers; viable system The organization contains operating units, feedback channels, coordination, control, and adaptation at several levels. An organization is also a normative and political order, not merely a regulated dynamical system.
Role Control objective, policy allocation, or supervisory mode A role specifies which decisions belong to an actor and which goals or constraints it should regulate. Roles also carry duties, legitimacy, status, and responsibility, which have no direct control-theoretic equivalent.
Agent Controller or decision-making subsystem The agent selects actions in response to goals, observations, and state. A persistent identity and task history are not required properties of a mathematical controller.
Harness Controller implementation, observer, planner, and feedback architecture The model alone is not the controller. State estimation, memory, tools, verification, and stopping logic determine the effective policy. A harness may also contain social instructions and interfaces that exceed a classical controller.
Habitat Admissible state-action region, constraint set, viability domain The habitat determines which resources, actions, and effects are available while the loop operates. A habitat includes institutional legitimacy and authority, not only physical or mathematical feasibility.
Sandbox Hard state constraints, actuator limits, isolation boundary Some transitions are made physically impossible regardless of the controller’s policy. A sandbox does not express the complete organizational meaning of an action.
Capability Enabled control input or admissible transition A capability adds a specific action to the controller’s available input set. Capabilities can be purpose-bound, delegated, expiring, and institutionally issued.
Tool Actuator, sensor, or combined transducer Tools translate decisions into effects and return observations about those effects. Some tools also perform substantial reasoning and may themselves contain controllers.
Workspace Local plant state or controller-accessible state The workspace contains mutable state available during the task. It may also contain records with ownership, confidentiality, and retention requirements.
Handoff policy Mode transition, switching law, guard condition, and reset map A handoff controls when authority and state move between actors or operating modes. Institutional handoffs also transfer reliance, responsibility, and legitimacy.
Approval Supervisory enable signal or transition guard Approval can unlock a previously disabled action or mode. A meaningful approval is a judgment supported by evidence, not merely a Boolean input.
Control plane Supervisory controller, governor, or metacontroller It configures lower-level controllers, constraints, modes, and authority without performing every local action. It must also reconcile independent sources of organizational authority.
Evidence Observation history, trace, residual, or state estimate support Evidence makes it possible to reconstruct what happened and evaluate the controller’s decision. Evidence has semantic and institutional meaning beyond sensor data.
Attestation Certified observation or acceptance of a transition condition An attestation states that a condition was checked before a transition proceeds. It binds an accountable party to the claim; control theory has no native concept of answerability.
Escalation Controller switching or transfer to a higher-level supervisor The system changes controllers when the local one lacks authority, information, or response variety. Escalation may be political, legal, or professional rather than technically superior control.
Post hoc scrutiny Offline evaluation, system identification, and outer-loop adaptation Traces from individual runs update future policies, models, constraints, and interfaces. Institutional scrutiny also assigns reasons and responsibility for past action.
Accountability No direct equivalent It determines who must explain and bear the consequences of an action. Causal control is not moral, legal, or institutional responsibility.

Habitat and the Reachable Set

The reachable set is the set of states a system can reach from a given initial state under its available inputs and dynamics over some time horizon.

Informally:

ReachableSet(current state, time) =
    every future state the system can produce
    using the actions actually available to it

This is close to the idea of a habitat, but it is not identical.

A habitat defines the actions and effects the organization has made admissible. The reachable set describes what the actor can actually cause from where it currently stands. The latter depends on more than permission.

Actual reachable effects =
    current system state
  + plant dynamics
  + available actuators
  + agent policy
  + harness competence
  + available information
  + habitat constraints

An agent may be authorized to repair a service but lack the knowledge required to find the fault. The state is inside its habitat but outside its effective reachable set. Another agent may discover an unintended route to a prohibited effect through a misconfigured tool. The state is outside the declared habitat but inside the technically reachable set.

That mismatch is one of the most important things an agent platform must detect.

Declared habitat ≠ Technically reachable set

A sound system tries to ensure that prohibited organizational effects are not technically reachable, while legitimate task outcomes remain reachable enough for the agent to be useful.

The design problem is therefore not minimum reachability. It is aligned reachability.

Four Sets That Should Not Be Collapsed

It is helpful to distinguish four nested or overlapping sets:

The imaginable action set

Everything the agent can represent or propose. A model may be able to describe actions that no available tool can perform.

The tool-enabled action set

Everything the harness can attempt through its actuators. This is the practical control-input set exposed by tools and interfaces.

The technically reachable set

Everything the combined agent, harness, tools, and plant dynamics can actually cause from the current state. Composition and unintended paths matter here.

The institutionally admissible set

Everything the organization considers legitimate for this actor, role, task, and moment. This is the normative core of the habitat.

The desired relationship is not perfect equality. An agent can imagine prohibited actions without being able to perform them. Some legitimate outcomes may require a handoff rather than direct reachability. But the boundaries should be intentional:

directly reachable by agent
    ⊆ permitted inside current habitat

reachable through governed handoff
    ⊆ permitted by the wider organization

prohibited organizational effects
    ∩ technically reachable effects
    = ∅, ideally

The final condition is an aspiration rather than a guarantee. Complex systems rarely have perfectly known dynamics. Defense therefore also requires observation, recovery, and adaptation.

Controllability Is Not Authority

In control theory, controllability asks whether available inputs can drive the plant from one state to another. It is a property of the system’s dynamics and actuators.

Authority asks a different question: whether this actor is entitled to produce that transition.

An agent may be controllable in the mathematical sense but unauthorized in the institutional sense. It may possess an actuator capable of deleting a database without possessing legitimate authority to use it. Conversely, it may have authority to achieve an outcome but lack the tools, information, or competence required to make that outcome reachable.

Controllability = can this transition be caused?
Authority       = may this actor cause it?
Competence      = can this actor find and execute a suitable policy?
Accountability  = who must answer for the resulting transition?

This four-way distinction prevents several common confusions. Giving an agent a tool increases controllability. Giving it a capability increases authorized control. Improving the harness may increase competence. Assigning a responsible role creates a place for accountability.

None substitutes for the others.

Observability Is Not Logging Everything

In control theory, a system is observable when its internal state can be inferred from the outputs available over time. Observability is not the same as producing a large quantity of telemetry.

An agent platform may log every tool call while remaining institutionally unobservable. If the trace cannot show which customer was affected, which policy applied, what state changed, or why the transition was allowed, the important state cannot be reconstructed.

Conversely, a compact evidence artifact may provide greater observability than a complete transcript. A release decision may be understandable through the changed interfaces, test results, known risks, responsible role, and rollback plan. The thousands of tokens used to produce it may add little.

The relevant question is:

Can the organization infer the state that matters from the evidence the system preserves?

Institutional observability is always relative to a question. Debugging, security review, legal accountability, and performance evaluation require different projections of the same underlying trace.

The Habitat as a Viability Constraint

Cybernetics and viability theory ask whether a system can continue operating without leaving a set of acceptable states. The viability kernel is the set of starting states from which there exists at least one policy that can keep the system inside those constraints.

This gives the habitat a second interpretation. A habitat should not only prevent bad effects. It should leave enough viable paths for useful work.

A habitat that blocks every dangerous action but also makes the task impossible is safe only in a trivial sense. A habitat that permits every action maximizes short-term controllability while destroying governance. The useful habitat preserves a region in which the agent can adapt, recover, and complete its role without crossing institutional boundaries.

This is the control-theoretic version of structured freedom.

The boundary must be tight enough to preserve the system and wide enough to contain a successful policy.

Handoffs as Hybrid-System Transitions

Agentic organizations rarely operate under one controller or one policy. They switch between agents, humans, automated pipelines, review processes, and emergency procedures. This resembles a hybrid system: continuous or iterative activity inside a mode, combined with discrete transitions between modes.

research mode
    │ evidence complete
    ▼
review mode
    │ attestation accepted
    ▼
publication mode

The handoff policy acts like a transition guard. It determines whether the system may leave one mode and enter another. The handoff may also apply a reset: revoke one habitat, provision another, change the responsible role, freeze an artifact, or create a durable record.

This analogy helps explain why approving each internal command is the wrong level of control. Supervisory systems usually govern mode transitions and operating envelopes. They do not manually select every low-level actuator movement.

Inside the mode, the local controller needs enough freedom to work. At the transition, the larger system checks whether the conditions for a new kind of consequence have been satisfied.

Reversibility, Recoverability, and Safe Return

The habitat framework uses reversibility to decide how much autonomy an action should receive. Control theory offers several more precise nearby concepts.

  • Backward reachability asks which earlier states could lead into a target set, including an unsafe set.
  • Recoverability asks whether the system can return to an acceptable region after a disturbance or error.
  • Region of attraction describes the starting states from which a controller will converge toward a desired state.
  • Invariant set describes a region the system can remain inside once entered.

Institutional reversibility is broader than any one of these. Restoring a database may recover technical state without recovering confidentiality, trust, or legal standing. A retracted message may remain in the recipient’s memory. Compensation may repair harm without undoing the original action.

Still, the control vocabulary gives a better question than simply asking whether an action has an undo button:

After this action, does the organization retain a credible path back to an acceptable state?

The smaller that recovery set becomes, the stronger the case for prevention, independent review, or formal risk acceptance before the transition.

Requisite Variety and Escalation

Ashby’s law of requisite variety says, roughly, that a regulator needs enough response variety to handle the variety of disturbances it faces.

For an agentic organization, this means a local agent cannot be expected to handle every exception merely because it is capable of general reasoning. It may lack information, tools, authority, or a valid response for the situation. A habitat deliberately limits its control variety.

Escalation supplies additional variety from elsewhere in the system. A security role can respond to a credential incident. A legal role can interpret an obligation. A human can recognize a political or relational consequence absent from the agent’s model. Another agent may contribute specialized competence without receiving broader authority.

An escalation path is therefore not evidence that autonomy failed. It is part of the architecture that makes bounded autonomy viable.

No local controller needs to contain the variety of the whole organization if it can recognize when to transfer control.

Inner Loops, Outer Loops, and Adaptation

The agent performs the task inside an inner loop. It observes, acts, checks, and adjusts. The harness determines the quality of this loop.

Around it sits an outer loop. The organization examines traces, handoffs, outcomes, incidents, and recurring failure patterns. It changes future role definitions, harnesses, habitats, tools, and policies.

Inner loop:
    improve the current action

Outer loop:
    improve the system that produces actions

This is a cybernetic account of institutional learning. A bad result is not only an error to correct. It is information about the controller, the environment, or the boundary. If the same class of failure recurs, the problem may no longer belong to the individual agent. It may be a defect in the harness, a missing sensor, an unreachable goal, an ambiguous handoff, or a habitat that grants the wrong control surface.

The agent learns inside the habitat. The organization learns by changing the next one.

Where the Analogy Breaks

Control theory is a language of causation, constraint, and feedback. Organizations are also systems of meaning, legitimacy, conflict, and responsibility.

A controller does not need a right to control the plant. A state transition does not need to be justified to the state it changes. A stable system is not necessarily a fair one. A highly observable organization may be an intolerable surveillance system. A perfectly enforced objective may be the wrong objective.

The most important institutional terms therefore have no complete control-theoretic equivalents:

  • Authority is legitimate permission, not mere access to an actuator.
  • Responsibility is an obligation attached to a role, not causal contribution alone.
  • Accountability is a relationship in which an actor must provide reasons to a forum capable of judgment and consequence.
  • Legitimacy concerns why the structure of control should be accepted at all.

Cybernetics can help us design how action is bounded, observed, corrected, and distributed. It cannot decide by itself who ought to control whom or which ends an organization should pursue.

The analogy is most useful when it makes the machinery visible without pretending the machinery settles the politics.

Compressed Vocabulary

Role          = allocation of institutional decision rights
Agent         = persistent decision-making actor
Harness       = controller and feedback architecture
Habitat       = institutionally admissible state-action region
Sandbox       = hard execution constraint
Capability    = enabled and authorized control input
Tool          = sensor, actuator, or both
Handoff       = governed mode and authority transition
Control plane = supervisory controller and policy governor
Evidence      = semantically useful observation history
Attestation   = accountable acceptance of a condition
Escalation    = transfer to a controller with different variety or authority
Audit         = retrospective state reconstruction and evaluation

The habitat is not simply the reachable set. It is the organization’s claim about which part of that set should be available to this actor, in this role, for this task.

The design objective is not to eliminate reach. It is to align technical reachability, institutional authority, and recoverable consequence.