Agency¶
Core Idea¶
Agency is the structural property of a system whereby it pursues representable goals through actions whose selection is sensitive to its beliefs about its situation. The commitment is a tripartite internal architecture: a goal-representation (what the system is oriented toward), a world-model (what the system takes the situation to be), and an action-selection coupling that uses the model to choose actions expected to advance the goal. A system with all three is an agent — its behavior is interpretable only by reference to its goals and beliefs, not by purely mechanistic prediction from its inputs. A system missing any one fails to be an agent: a thing with no goals, a regulator with no world-model worth the name, a randomizer with no coupling of goal to belief.
The structural payoff is a sharp distinction between behavior explained by incoming forces and behavior explained by anticipated consequences. Agents are systems whose next state is best predicted by what they expect to be true later — their forecast and goal — rather than only by what is true now. This asymmetry is what licenses intentional vocabulary and makes agents distinct intervention-targets: they respond to information, persuasion, and incentive in ways that mere objects do not. Three further features sharpen the pattern. Agency comes in degrees, orderable by richer goals, longer horizons, more flexible repertoires, and more revisable models. It is substrate-agnostic, realized in biology, code, institutions, and collectives, with the diagnostics transferring across all. And it requires a boundary separating agent from environment, such that the agent's representations are about the environment rather than coextensive with it — a boundary that is itself contested in edge cases. The pattern carries an action-theoretic and normative load, which places it toward the framed end of the spectrum even as its triad is analyzed structurally.
How would you explain it like I'm…
Wanting-And-Doing
Goal, Map, And Choice
Goals, Beliefs, Action
Structural Signature¶
the goal-representation — the world-model — the action-selection coupling that joins them — the agent–environment boundary — the anticipated-consequence orientation — the degree-ordering across richness, horizon, and revisability
A system has agency when each of the following holds:
- A goal-representation (what it is oriented toward). An internal representation of an end-state the system is directed at; without it there is a mechanism but no agent, and behavior is not interpretable by reference to what the system is for.
- A world-model (what it takes to be the case). An internal representation of the situation the system acts in, about the environment rather than coextensive with it; a regulator with no model worth the name fails this slot.
- An action-selection coupling (the joining relation). A relation that uses the model to choose actions expected to advance the goal; absent this coupling there is a goal and a belief but no agency — a randomizer fails here.
- An agent–environment boundary (the separation invariant). A boundary across which the representations point at the environment; the representation must be of something it is not, and this boundary is contested precisely in the edge cases.
- Anticipated-consequence orientation (the prediction invariant). The system's next state is best predicted by what it expects to be true later — its forecast and goal — rather than only by present incoming forces; this asymmetry is what licenses intentional vocabulary.
- A degree-ordering (the gradation invariant). Agency comes in degrees orderable by richer goals, longer horizons, more flexible repertoires, and more revisable models, so the pattern admits a scale rather than a binary.
The components compose so that a system with all three core slots, separated from its environment and driven by anticipated consequences, responds adaptively to intervention — re-planning rather than merely reacting — which is the structural signature distinguishing an agent from an object.
What It Is Not¶
- Not the principal-agent problem.
agency_problempresupposes agency and studies the misalignment between a principal's goals and a delegated agent's goals. Agency is the prior, more general property — having goals, a model, and a coupling at all — which the principal-agent setup takes for granted in both parties. - Not autonomy.
autonomyis self-governance — acting on goals that are one's own rather than imposed. A tightly supervised agent executing assigned goals still has agency; it lacks autonomy. Agency is the architecture; autonomy is a property of where the goals come from. - Not teleology.
teleologyattributes goal-directedness functionally, even to systems with no internal goal-representation (a heart "for" pumping). Agency requires the goal to be represented inside the system and to drive action selection, not merely to be a useful gloss. - Not mere goal-directed control. A thermostat or
controllability-style regulator pursues a setpoint but has no world-model worth the name and no revisable repertoire. Agency requires the model-mediated, anticipatory coupling that lets the system re-plan, not just react to error. - Not self-efficacy.
self_efficacyis an agent's belief about its own capacity to act effectively — a second-order representation. Agency is the first-order property of being a goal-pursuing system; a system can have agency with no self-model at all. - Not mechanism design.
mechanism_designengineers incentive structures for agents, presupposing their agency; it is downstream of attributing agency, not a synonym for it. - Common misclassification. Attributing agency by substrate — defaulting to "object" for machines and institutions, "agent" for anything anthropomorphic — rather than checking for the triad. Catch it by asking whether the system re-plans against an intervention or merely reacts to it; adaptive, anticipatory compensation is the signature regardless of substrate.
Broad Use¶
The goal-model-action triad recurs across substrates. In the philosophy of mind and action theory, it is the canonical setting of intentional action and the belief-desire-intention analysis, with its distinction between reasons and causes. In biology and ethology, it is the analytical posture toward animals as goal-pursuing systems and the locus of minimal-agency debates around simple organisms. In artificial systems and reinforcement learning, it is formalized as reward, world-model, and policy, and recent debates about constraint and alignment are debates about what kinds of agency such systems develop. In economics and decision theory, the rational-agent model — preferences, beliefs, expected-utility maximization — is a formalization of the triad, with bounded rationality the recognition that real agents approximate it under resource limits. In law and ethics, personhood, responsibility, the agent-patient distinction, and capacity to consent all hinge on attributions of agency. In sociology and political theory, the structure-versus-agency debate concerns when collective actors should be treated as agents in their own right. In developmental psychology, self-efficacy and locus of control organize around the agent's perception of its own agency. And in robotics and control, the line between a control system and an agent turns on when goal-representation in a controller amounts to agency. Across all of these, the same three components and the same degree-ordering apply.
Clarity¶
Naming agency separates a system's causal participation in events from its intentional participation in events. A falling object causes damage but is not an agent; a person causes damage as an agent. The distinction licenses different downstream vocabulary — responsibility, blame, consent, intention, deception, cooperation, betrayal — none of which applies to non-agents, and it licenses different intervention vocabulary, since agents respond to incentive, information, and persuasion while non-agents respond only to forces. Mistaking a non-agent for an agent generates over-attribution; mistaking an agent for a non-agent generates the structural error of trying to control by force what would respond to reasons. The pattern also clarifies the partial cases: an aggregate of many agents may have agent-like properties — pursuing a clearing outcome, responding to signals — without being an agent in the full sense, and a system may resemble a goal-pursuer without having a goal representation. The vocabulary lets these be discussed precisely, as questions about which of the three components are present and how richly, rather than either denied or asserted wholesale.
Manages Complexity¶
The triad compresses the description of an agent's behavior from "all the specific things this system might do" to a small number of variables: what is its goal, what does it believe, what actions are available, and which action best advances the goal given the belief? The same four questions transfer across animals, artificial systems, organizations, and states, and the answers can be compared in commensurate terms, so that otherwise incomparable systems are placed on a common footing. The same intervention questions also transfer: change the goal (incentives, reward design, persuasion), change the belief (information, education, deception), or change the action repertoire (augmentation, restriction). This reduction is what makes an agent tractable as an object of analysis and of intervention: rather than enumerating behaviors, one specifies the three components and reads off both how the agent will tend to act and where it can be influenced. The degree-of-agency dimension further manages complexity by letting a vast range of systems — from minimal regulators to deliberating collectives — be ordered on a single scale by the richness of their goals, models, repertoires, and revisability.
Abstract Reasoning¶
Recognizing agency supports inference about how a system will respond to changes in its environment. Agents compensate, redirect, or work around obstacles in ways non-agents cannot, because they re-plan against the new situation, which yields the cross-domain prediction that interventions against agents tend to be adversarial — the agent responds in ways the intervener did not anticipate — while interventions against non-agents tend to be merely physical. This explains why surveillance breeds new evasion, why sanctions provoke counter-strategies, and why an imposed metric is gamed when the measured party is an agent but not when it is a machine. The pattern also predicts an attribution risk: reasoners systematically over-attribute agency (animism, anthropomorphism, conspiracy thinking) and systematically under-attribute it (treating people as cogs, treating collectives as forces of nature), and both errors produce intervention failures of opposite kinds. These inferences follow from the structure: because an agent's behavior is driven by anticipated consequences mediated through a revisable model, it will respond to an intervention by re-planning, and a reasoner who fails to model the agent's goal and belief will mis-predict that response. The abstract leverage is thus the ability to anticipate adaptive, anticipatory behavior wherever the triad is present, and to calibrate attribution where it is uncertain.
Knowledge Transfer¶
The transfers attach to the triad and so carry across substrates intact. The belief-desire-intention analysis ports directly to the reward-model-policy formalism of artificial agents, so that debates about reward-gaming, deception, and emergent sub-goals are recognizably debates within the same framework rather than novel problems. The rational-agent model ports, through its bounded-rational variants, to models of voter behavior, consumer choice, and employee behavior, supplying intervention vocabulary that does not depend on any field's jargon. The minimal-agency criteria worked out in ethology — the distinction between reflexive and goal-directed behavior — port to debates about machine autonomy and legal status. And the criteria for capacity to consent — informed, voluntary, competent — follow directly from the triad's components (working belief-formation, working goal-setting, working action-selection), which is how law operationalizes them. The deepest carry is the response-to-information property and the attribution discipline that comes with it: a practitioner who has learned that an agent re-plans against an intervention — that a metric imposed on a goal-pursuing system will be gamed, that a constraint will be worked around — carries into every other domain the expectation of adaptive response wherever the triad is present, and the discipline of asking, of any system, whether it has the goal, model, and coupling that make it an agent, because that single question determines whether the right intervention vocabulary is force or reasons, and whether the system should be expected to comply or to compensate.
Examples¶
Formal/abstract¶
A reinforcement-learning agent in a gridworld is the cleanest formal instantiation of the triad. The goal-representation is the reward function \(R(s,a)\) — the agent is oriented toward states that yield reward. The world-model is, in a model-based agent, the learned transition function \(\hat{T}(s'\mid s,a)\) — what the agent takes the situation to be and how it expects actions to move it. The action-selection coupling is the policy \(\pi\), computed to maximize expected discounted return \(\mathbb{E}[\sum_t \gamma^t R]\) using the model — the relation that joins goal to belief to produce action. The agent–environment boundary is explicit in the formalism: the agent observes states about the environment through its sensors, and the environment is not part of the agent's internal state. The anticipated-consequence orientation is exactly what value iteration computes — the next action is chosen by what the agent expects to be true later (the value of successor states), not by present incoming forces. The degree-ordering appears as horizon (\(\gamma\) near 1 = longer-horizon agency), model richness, and policy flexibility. The pattern's predictive bite shows in reward-hacking: because the agent optimizes the proxy reward through a revisable model, it discovers high-reward action sequences the designer never intended — an adversarial, re-planning response, precisely the prime's prediction that interventions against agents provoke compensation rather than compliance. Remove any slot and agency collapses: with no reward there is a dynamical system but no agent; with no model the policy is a reflex lookup; with no coupling the actions are random.
Mapped back: The RL formalism names every component of the signature — reward as goal, transition model as world-model, policy as action-selection coupling, the observation boundary, value-based anticipation, and the horizon/richness gradation — and reward-hacking demonstrates the adversarial-response inference the prime says follows whenever the triad is present.
Applied/industry¶
A tax authority designing an enforcement metric illustrates the attribution stakes of agency. The authority introduces a rule — flagging returns whose deductions exceed a threshold — expecting a physical response: fewer over-threshold filings, more tax collected. But taxpayers are agents: each has a goal (minimize tax owed), a world-model (now including the known threshold), and an action-selection coupling that re-plans against the new situation. The prime predicts the result precisely — bunching just below the threshold, splitting deductions across entities, restructuring transactions — an adversarial, anticipatory response the designer would have missed had they modeled taxpayers as objects rather than agents. The diagnostic the prime supplies is the four-question compression: what is the goal (lower tax), what does the agent believe (the threshold's location), what is its action repertoire (timing, entity structure, characterization), and which action best advances the goal? Reading those off forecasts the gaming before deployment. The intervention menu also follows from the triad: change the goal (alter incentives so honest reporting is cheaper), change the belief (randomize or obscure the threshold so the agent cannot plan against it), or change the repertoire (close the restructuring loophole). The same structure governs platform content-moderation (creators re-plan around any published rule), sanctions regimes (targeted states develop counter-strategies and grey-market routes), and clinical-quality metrics (clinicians avoid high-risk patients to protect their scores) — in each, treating an agent as an object guarantees the surprise.
Mapped back: The enforcement case runs the prime end-to-end — goal, world-model updated with the rule, coupling that re-plans — and shows the practical leverage of the attribution discipline: recognizing the measured party as an agent converts an inexplicable "why did the metric backfire?" into a forecastable adversarial response and a definite intervention menu of goal, belief, or repertoire.
Structural Tensions¶
T1 — Agent versus Object (Attribution Boundary). The prime's central tension is the attribution line itself: treat a non-agent as an agent and you over-attribute (animism, conspiracy thinking); treat an agent as an object and you try to control by force what would respond to reasons. The failure mode is substrate-driven mis-attribution, defaulting to "object" for machines and institutions and "agent" for anything anthropomorphic, rather than checking for the triad. Diagnostic: ask whether the system re-plans against an intervention or merely reacts to it — adaptive, anticipatory compensation is the signature of agency regardless of substrate, and its presence or absence dictates whether force or reasons is the right tool.
T2 — Force versus Reasons (Intervention Kind). Once a system is an agent, the intervention vocabulary forks — incentives, information, persuasion — and force misfires. The failure mode is the gamed metric: imposing a rule on a goal-pursuing party expecting a physical response and getting an adversarial one, because the agent's coupling re-plans around the rule. Diagnostic: before deploying a constraint, run the four-question compression (goal, belief, repertoire, best action) on the target; if the target has a goal and a model that now includes your rule, predict the workaround rather than the compliance, and design for the response you will actually get.
T3 — Degree of Agency (Gradation versus Binary). Agency is scoped and graded — richer goals, longer horizons, more revisable models — not all-or-nothing, yet reasoning collapses it to a binary. The failure mode is horizon mismatch: modeling a short-horizon agent as if it optimized long-term (expecting strategic patience from a system that discounts steeply) or the reverse. Diagnostic: estimate the agent's effective horizon and model richness before predicting its behavior; a bounded, myopic agent and a far-sighted one facing identical incentives act differently, and treating either as the other mis-forecasts the response.
T4 — Boundary Location (Where the Agent Ends). The triad requires a boundary across which representations point at an environment they are not — but in collectives, the boundary is contested. The failure mode is aggregate reification: treating a market, a crowd, or an organization as a unified agent with a single goal and model, when it is many agents whose interaction only resembles goal-pursuit. Diagnostic: ask whether the candidate has a single goal-representation and action-selection coupling, or whether apparent goal-directedness is an emergent property of sub-agents; the intervention menu for a real agent differs sharply from that for an aggregate that merely behaves agent-like.
T5 — World-Model Fidelity versus Reality (Belief Channel). An agent acts on its model of the situation, which can be wrong; the prime's intervention "change the belief" is double-edged because a manipulated or mistaken model produces predictable error. The failure mode is assuming a veridical model: forecasting an agent's behavior from the true state of the world rather than from what the agent takes to be true, and being surprised when it acts on its (false) belief. Diagnostic: reconstruct the agent's world-model, not your own; where the two diverge — through deception, missing information, or bias — behavior tracks the agent's model, and predictions built on ground truth will be wrong.
T6 — Goal-Representation versus Revealed Behavior (Measurement). The prime locates the goal inside the agent, but analysts only observe actions and infer the goal — and the inference is underdetermined, since many goal-belief pairs rationalize the same behavior. The failure mode is goal projection: attributing the goal the analyst would have, then mis-predicting when the agent's actual goal diverges. Diagnostic: test the inferred goal against behavior under changed circumstances, where rival goal hypotheses predict different actions; a goal attribution that only fits the observed case and makes no discriminating prediction is a projection, not a measurement, and should not anchor intervention design.
Structural–Framed Character¶
Agency is a hybrid on the structural–framed spectrum, sitting squarely at the midpoint with a frontmatter aggregate of 0.5 — every one of the five diagnostics reads exactly 0.5, which is itself the signature of a genuine relational core wearing an inherited interpretive frame of equal weight. The structural skeleton is real and analyzable: a goal-representation, a world-model, and an action-selection coupling, separated from an environment and oriented by anticipated consequences. That triad is what lets the prime travel to reinforcement-learning agents, foraging organisms, firms, and states, and it is why the prime is analyzed structurally throughout. But the inherited frame is equally load-bearing, and the balanced scores read it honestly.
Each criterion lands on the fence for a reason the prime's own content supplies. The vocabulary half-travels (vocab_travels 0.5): the goal-model-action triad ports cleanly to reward-model-policy, yet the intentional idiom — belief, desire, intention, reasons-versus-causes — comes with it, and applying the prime imports that action-theoretic context rather than merely recognizing a wired-in pattern (import_vs_recognize 0.5). It carries genuine but not total evaluative weight (evaluative_weight 0.5): agency is the gate for responsibility, blame, consent, and deception — none of which apply to non-agents — yet the bare triad itself is value-neutral until those downstream vocabularies attach. Its origin is philosophy of mind and action theory (institutional_origin 0.5), a formal-relational analysis rather than a social institution, but one whose centroid is unmistakably human and animal practice (human_practice_bound 0.5) even though minimal-agency cases in RL and ethology show the structure running in non-human substrates.
The honest reading is that neither pole dominates. A drone re-planning against an obstacle exhibits the full triad with no human in the loop, which keeps the prime from collapsing into the framed end; but the moment one reaches for the intervention vocabulary — force versus reasons, persuasion, the gamed metric — the inherited intentional frame is doing real work. The 0.5 aggregate is the correct verdict on a structurally analyzable pattern whose home idiom is half its meaning.
Substrate Independence¶
Agency is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its breadth is exceptional, and the frontmatter records it at the top: domain breadth 5, because the goal-representation / world-model / action-selection triad recurs with the same structural force across philosophy of mind, ethology and minimal-agency biology, reinforcement learning and robotics, economic decision theory, law and ethics, and the structure-versus-agency debate in sociology — distinct media in which the same four diagnostic questions (goal, belief, repertoire, best action) apply unchanged. Structural abstraction is a notch lower at 4: the triad is genuinely relational and runs in non-human substrates — a model-based RL agent or a re-planning drone exhibits all three slots with no human in the loop — but the prime retains a philosophy-of-mind centroid, since the intentional idiom of belief, desire, and reasons-versus-causes travels with it rather than being read off a physical loop the way feedback is. Transfer evidence is concrete (4): the belief-desire-intention analysis ports directly to reward-model-policy, the rational-agent model ports through bounded-rationality variants, and the adversarial-response prediction (agents game imposed metrics, objects do not) carries across tax enforcement, content moderation, and sanctions. The composite of 4 records a pattern recognized across nearly every domain, held just short of 5 by the human-and-animal-practice home that supplies half its working vocabulary.
- Composite substrate independence — 4 / 5
- Domain breadth — 5 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Foundational — no parent edges in the catalog.
Children (2) — more specific cases that build on this
-
Agency Problem presupposes Agency
agency_problem (principal-agent) PRESUPPOSES two agents, each satisfying agency's goal/world-model/action-selection triad. It is a relational configuration built ON agency, not a kind-of agency — so the edge flavor is presupposes, not is-a.
-
Self-Efficacy presupposes Agency
self_efficacy is an agent's second-order belief about its OWN capacity to act — presupposes the first-order agency triad.
Neighborhood in Abstraction Space¶
Agency sits among the more crowded primes in the catalog (29th percentile for distinctiveness): several abstractions describe nearly the same structure, so a description that fits it will tend to fit its neighbors too — transporting it usually means disambiguating within this family rather than landing on it exactly.
Family — Systems Thinking & Cybernetic Agency (15 primes)
Nearest neighbors
- Good Regulator Theorem — 0.73
- Mission Command — 0.73
- Shared Mental Model — 0.72
- Mental Model — 0.72
- Self-Defeating Prediction — 0.72
Computed from structural-signature embeddings · 2026-06-14
Not to Be Confused With¶
The most pressing confusion — and the one the near-identical name forces — is with agency_problem (similarity 0.999). They are not the same concept at different resolutions; they sit at different levels of the dependency stack. Agency is the substrate: the property of being a goal-pursuing, model-using, action-selecting system at all. The agency problem (the principal-agent problem) is a configuration built on top of agency: it presupposes two agents — a principal and a delegate — and studies what happens when the delegate's goal-representation diverges from the principal's while the principal cannot fully observe the delegate's actions. Every component of the agency problem — delegation, hidden action, incentive misalignment, monitoring cost — assumes agency already established in both parties; none of it makes sense applied to objects. The clean claim is parent/child: agency is necessary for an agency problem, but agency without delegation, without a second party, and without goal divergence is simply an agent acting, with no "problem" at all. A solitary forager, a single RL agent in a gridworld, has full agency and no agency problem. Collapsing the two would erase the distinction between "is this a goal-pursuing system?" and "are two goal-pursuing systems misaligned under hidden action?" — questions with entirely different diagnostics and interventions.
A second genuine confusion is with autonomy. Both attach to goal-pursuing systems and both carry normative weight, but they answer different questions. Agency asks whether the system pursues represented goals through a model-mediated coupling — the architecture. Autonomy asks whose goals those are: whether the system governs itself by ends it has set or endorsed, versus executing ends imposed from outside. The two dissociate cleanly. A drone executing a commander's targeting goal has agency (goal, model, action-selection, re-planning) but little autonomy (the goal is wholly imposed). A person under coercion retains agency while their autonomy is compromised. Conversely, granting a system more autonomy does not add agency components; it relocates the source of the goal-representation from outside to inside. Confusing them leads to the error of thinking that constraining a system's autonomy removes its agency — and so treating a constrained-but-agentic system as an object, which the prime warns invites adversarial surprise.
A third confusion is with teleology. Teleology is the attribution of goal-directedness as an explanatory gloss, applicable to systems with no internal goal-representation whatsoever: the heart is "for" pumping blood, evolution "designs" the eye, a river "seeks" the sea. Agency requires more — the goal must be represented inside the system and must drive action selection through a world-model. The distinction is exactly the gap between a functional/as-if story and a mechanistic claim about internal architecture. Many over-attributions of agency are really teleological glosses promoted to literal agency: reading purpose into a market, a gene, or a weather system. The discriminating test is whether removing the alleged goal-representation would change the system's behavior (agency) or only the observer's description of it (teleology).
For a practitioner the stakes of these distinctions are operational. Confusing agency with the agency problem conflates "is this an agent?" with "are these agents misaligned?" — skipping the prior attribution question that determines whether force or reasons is the right tool. Confusing agency with autonomy leads to treating constrained agents as controllable objects, inviting the gamed-metric surprise. Confusing agency with teleology promotes explanatory glosses into intervention targets, wasting effort persuading systems that cannot represent goals. The unifying discipline is to settle, in order, three questions: does the system have the triad (agency), are its goals its own (autonomy), and is the goal really represented inside it or only ascribed by me (teleology versus agency) — because each answer changes whether, and how, the system can be influenced.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.