Agentic Control Loop Design

Agency becomes real when goals, situation models, available actions, authority, execution, feedback, and learning are coupled into a loop that can intentionally change outcomes.

Essence

Agentic Control Loop Design makes agency structurally possible. It does not equate agency with autonomy, motivation, confidence, or authority alone. An agent needs a represented goal, a model of its situation, feasible actions, a way to select among them, permission and capability to execute, feedback about effects, and a rule for updating the next action.

When those links are coupled, action can become intentional and adaptive. When they are broken, “agency” becomes a slogan: people are blamed for outcomes they cannot control, teams wait for permission, software agents act without boundaries, or organizations mistake dashboards for empowerment.

Compression statement

Agentic Control Loop Design applies when a person, team, institution, software agent, or other system is expected to pursue goals but lacks a coherent coupling among goal representation, world model, action-option generation, action selection, permission to act, execution resources, feedback, and model update. The archetype builds or repairs that loop so action is not random motion, compliance, helpless reaction, or delegated execution without understanding, but belief-sensitive pursuit of represented goals under constraints.

Canonical formula: agency = represented_goal + world_model + action_options + selection_policy + execution_authority + feedback_update; design = couple_and_calibrate_these_links_under_constraints.

When to use it

Use this archetype when a person, team, institution, or software system is expected to pursue goals but the practical loop from goal to action to feedback is incomplete. It is especially useful when the environment is changing, local judgment matters, and centralized instructions cannot anticipate every condition.

The diagnostic question is simple: can the actor observe enough, understand enough, choose among real actions, act legitimately, see the effects, and update future choices?

Key components

Component	Description
Represented Goal ↗	Agency begins with a goal the actor can represent. The goal does not need to be perfectly precise, but it must be concrete enough to guide action and revisable enough to survive new evidence.
World Model ↗	The world model is the actor’s belief about the current situation, causal structure, constraints, and possible consequences. Without a model, action is not situation-sensitive. With a bad model, agency can become confident error.
Action Repertoire ↗	A system has agency only if it has feasible actions. Symbolic choice, motivational slogans, or outcome responsibility without action options create pseudo-agency.
Selection Policy ↗	The selection policy turns goals and beliefs into chosen actions. It may be a human judgment process, a rule, a heuristic, a protocol, an algorithm, or a learned policy.
Legitimate Action Boundary ↗	Agency needs room to act, but not unlimited discretion. The legitimate action boundary states what may be decided locally, what must be escalated, and what can be overridden.
Effect Feedback Loop ↗	Feedback connects action to consequence. Without it, the actor cannot learn whether its model or policy worked. Feedback should be timely, specific, and connected to action effects rather than only retrospective blame.
Proportional Accountability Frame ↗	Accountability should match actual control, information, and resources. This component prevents agency language from becoming a way to blame actors for structurally uncontrollable outcomes.

Common mechanisms

An agency loop map visualizes the full structure. A decision-rights matrix clarifies authority. Briefback or intent confirmation checks whether an actor understands the goal and constraints before acting. A controllability mapping checklist tests whether responsibility is fair. A graduated autonomy ramp expands discretion as evidence improves. After-action learning cycles convert outcomes into model updates. Safe action menus preserve bounded choice in high-risk settings.

Parameter dimensions

Important dimensions include goal clarity, model quality, observability, action range, resource sufficiency, legitimacy boundary, decision latency, feedback speed, error cost, autonomy level, support intensity, and accountability proportionality. The right design depends on whether the agent is a person, team, institution, or software system, and whether failure creates minor learning cost or serious harm.

Invariants to preserve

The actor’s goal should remain visible and contestable. Action authority should remain bounded. Feedback should be connected to action effects. Accountability should remain proportional to actual control. Model updates should be possible when evidence changes. Support should not become takeover.

Neighbor distinctions

Autonomous Action Zone Protection protects a legitimate decision space; this archetype builds the loop that makes action in that space meaningful. Self-Efficacy Scaffolding builds confidence; this archetype requires real control and feedback as well. Decision Rights Clarification assigns authority; this archetype also requires goals, models, action options, execution, learning, and accountability. Principal–Agent Alignment addresses delegated incentive conflict; Agentic Control Loop Design can apply even when there is no principal-agent misalignment.

Failure modes and safeguards

Pseudo-agency appears when actors are made responsible without authority or resources. Unbounded discretion appears when agency is treated as freedom from constraints. Feedback-as-punishment appears when results are used for blame rather than learning. Goal capture appears when an empowered actor pursues the wrong objective. Over-scaffolding appears when support never transfers control. The safeguards are controllability mapping, legitimate boundaries, feedback redesign, goal review, graduated autonomy, and proportional accountability.

Examples

A field team receives commander intent, local authority, situation updates, and after-action feedback. A learner gets meaningful choices, visible effects, and progressively larger responsibility. A product team receives customer-outcome goals, budget, decision rights, user feedback, and learning cadence. A software agent receives an objective, state representation, permitted tools, confidence thresholds, feedback, and escalation rules.

Non-examples

A dashboard without action rights is not agency. Motivation without feasible action is not agency. Delegation without intent, constraints, and feedback is not agency. Responsibility without controllability is not agency; it is blame transfer.

Common Mechanisms

• Action-Effect Feedback Review
• After-Action Learning Cycle
• Agency Health Dashboard
• Agency Loop Map
• Briefback or Intent Confirmation
• Controllability Mapping Checklist
• Decision-Rights Matrix — Maps each class of decision to who may decide, approve, be consulted, or merely be informed — fixing the agent's authority before any single choice arises.
• Graduated Autonomy Ramp
• Model Assumption Register
• Safe Action Menu

Related Abstractions

Abstractions this archetype builds on — directly (a source ingredient) or as a related pattern. Links follow the typed catalog namespace.

Built directly on (6)

• Agency: A system pursues representable goals through actions whose selection is sensitive to its beliefs about its situation, via a goal-representation, world-model, and action-selection coupling.
• Controllability: Ability to steer system.
• Feedback: Outputs influence inputs.
• Good Regulator Theorem: Every effective regulator of a system must be, or must contain, a model of that system.
• Mental Model: Internal system representation.
• Perception Action Loop: Perception and action are constitutively coupled: action moves the sensing apparatus, that movement changes what is sensed, and what is sensed becomes the basis for the next action, in one closed loop with no clean sense-think-act stages.

Also references 20 related abstractions

• Accountability: Responsibility for actions.
• Agency Problem: Misaligned incentives.
• Autonomy: A unit's behavior is governed by its own internal rules or chosen reasons rather than external direction, defined by the inner-versus-outer authority asymmetry over a scoped domain.
• Boundary: Defines system limits.
• Bounded Rationality: Limited decision capacity.
• Collective Efficacy: Shared belief in capability.
• Constraint: Limits possibilities to guide outcomes.
• Decision: Committing to one alternative from a set under uncertainty and trade-off, collapsing open deliberation into a chosen path and foreclosing the others.
• Delegation of Authority: Assign responsibility.
• Governance: The durable architecture of authority, accountability, and decision rights through which a group makes binding collective choices and resolves disputes internally.

▸ Show 10 more

Variants

Narrower or domain-specific specializations that share this archetype's core structure. Recognized variants are established; candidate variants are provisional.

Human Agency Restoration · affective or cognitive variant · recognized

A variant focused on restoring real and perceived ability to act after passivity, learned helplessness, or disempowerment.

• Distinct from parent: Narrower and psychologically sensitive.
• Use when: The actor has experienced repeated uncontrollability; Small real choices and visible effects can rebuild action-effect coupling; Support must avoid takeover.
• Typical domains: education and coaching, healthcare and behavioral support, community development
• Common mechanisms: action effect feedback review, graduated autonomy ramp

Delegated Agentic Mandate · governance variant · recognized

A governance variant where a principal gives intent, boundaries, resources, and feedback so an agent can adapt locally.

• Distinct from parent: Narrower and more governance-oriented.
• Use when: Central instruction cannot keep up with local conditions; The agent needs bounded discretion; Accountability and escalation can be made explicit.
• Typical domains: mission command and operations, organizational management
• Common mechanisms: briefback or intent confirmation, decision rights matrix

Software Agent Control Loop · implementation variant · candidate

A software or AI-system variant where objective, state model, tool-use permissions, action policy, monitoring, and escalation are explicitly coupled.

• Distinct from parent: More technical and safety-sensitive than the general parent.
• Use when: A software agent or automation can act in an environment; Tool access, objective, model confidence, and escalation need boundaries; Feedback must update future action selection.
• Typical domains: software agents and ai systems, robotics and autonomous systems
• Common mechanisms: safe action menu, agency health dashboard

Collective Agency Coordination · governance variant · candidate

A group-level variant where shared goals, shared models, distributed action rights, and feedback allow a collective to act as a coherent agent.

• Distinct from parent: Adds coordination and shared-model requirements.
• Use when: Multiple actors need to pursue a shared goal; No single actor controls the whole system; Coordination, feedback, and role clarity are necessary for collective action.
• Typical domains: community development, incident response, organizational management
• Common mechanisms: briefback or intent confirmation, after action learning cycle

Near names: Goal-Model-Action Coupling, Agency Loop Design, Agentic Capacity Design, Belief-Sensitive Action Design.