When Intelligence Becomes Agency: A Theory of Governed, Proactive Agency for Symbiotic AI Systems

1 Introduction

• The Activation Problem is deciding whether behavior is warranted, when it should occur, and whether to act, ask, monitor, defer, escalate, or refrain.Current agents mainly determine what and how to do after external activation, whereas proactive agency evaluates ongoing observations against standing commitments.Source: §1 · pp. 1–2
• Agency is an organizational property in which perception, intent, affective-conative inference, constraint, and feedback organize behavior across time.These elements define the agenthood minimum; fuller agency additionally requires selecting whether, when, and how to act, including restraint and observable records.Source: §1 · pp. 2–3
• Symbiotic agency is delegated agency under a standing, revocable mandate, continuous coupling to the principal, calibrated condition inference, and bounded personalization.Unlike a bounded delegated task, symbiotic assistance acts, waits, asks, escalates, or refrains to augment rather than replace the principal’s judgment.Source: §1 · pp. 3–4
• The taxonomy first requires the agenthood minimum, then classifies agency by mandate: autonomous without one, delegated with one, and symbiotic when delegation persists with continuing coupling.Ownership and goal origin do not determine the form; the mandate does, while ownership separately supplies operational control and shutdown authority.Source: §1 · pp. 3–4

1.1 Thesis, scope and reader’s guide

• The paper’s thesis is that intelligence becomes agency when these elements organize behavior under an answerable mandate, with behavioral episodes as the unit of analysis.Activation weighs commitments, authority, uncertainty, and response modes rather than treating a trigger as sufficient for action.Source: §1.1 · pp. 4–5
• The theoretical contribution integrates commitment, principal-state inference, appraisal, interruption management, mandate governance, and constrained decision-making into a framework for classifying and evaluating governed agency.Its instruments include a signature method, evaluation framework, benchmark direction, and reference layer stack; it reports no empirical results or single implementation.Source: §1.1 · p. 5

2 A Genealogy of Agency as Organized Behavior

• The historical progression moves from pre-arranged automata and stimulus–response explanations toward purposive, feedback-regulated accounts of organized behavior.Tolman’s goal-directed, molar behavior and cognitive maps, followed by cybernetic feedback, established analysis beyond isolated movements or local reactions.Source: §2 · pp. 5–6

2.1 Symbolic control, situated agency and the limits of action selection

• Symbolic planning, BDI commitment, situated interaction, and multi-agent accounts each capture aspects of agency but leave unresolved how behavior becomes prioritized, inhibited, and mandate-sensitive [12, 13, 20, 21, 22, 23, 24, 27, 28, 29].A durable intention can remain dormant or be deferred, while situated responsiveness alone may produce continuous reaction without stronger organization.Source: §2.1 · pp. 6–7

2.2 The LLM implementation turn

• LLM agents expanded agency-like behavior from isolated responses to long-horizon trajectories involving reasoning, tool use, observation, revision, persistent state, and termination decisions [30] [32] [34] [35] [36] [37] [38] [39] [40] [41] [42].Their implementation advances make activation more visible because sustained execution still requires a theory of whether, when, and in what mode capabilities should become behavior.Source: §2.2 · pp. 7–8

2.3 From agent mechanisms to organized behavior

• Agency requires an integrated account that determines whether, when, and in what mode behavior occurs under commitments, context, authority, and feedback.The relevant unit is the behavioral episode: a temporally extended pattern of noticing, evaluating, selecting, inhibiting, sequencing, revising, and sometimes refraining.Source: §2.3 · pp. 8–9

3.1 Behavior, action and restraint

• Agency is expressed through organized behavioral episodes whose actions, timing, sequencing, inhibition, revision, and restraint are coordinated around aims and constraints.The agent is the behavioral loop enabled by components such as models, tools, memory, planners, and orchestration, not any single component.Source: §3.1 · pp. 9–10

3.2 Intent as the organizing principle of behavior

• Intent makes behavior agentic by sustaining commitments that guide action selection across time and update through feedback.Intent distinguishes commitments from mere desires and organizes behavior beyond the number or sophistication of actions a system can perform.Source: §3.2 · p. 10

3.3 Affective-conative inference and the activation problem

• Affective-conative inference determines when commitments become active by appraising context, memory, user state, risks, permissions, priorities, and possible outcomes.Its outputs may include acting, asking, remembering, monitoring, delaying, escalating, or refraining; proactivity therefore means independently judging whether behavior is warranted, not always intervening.Source: §3.3 · p. 11

3.4 The sensorium: exteroception, agent-interoception, allo-interoception

• The sensorium combines exteroception, agent-interoception, and uncertain allo-interoception, while requiring authorized, relevant evidence because the agent cannot directly access the principal’s internal experience.Wider sensing can improve helpfulness but also heightens intrusion, surveillance, and third-party exposure, so intimate signals require consent, minimization, policy, explainability, and user control.Source: §3.4 · pp. 12–13

3.5 Affective-conative profiles

• Affective-conative profiles translate signals into priorities and behavioral orientations through generic, personalized, or evolvable mappings bounded by agency form, mandate, consent, correction, and policy.The personalized profile is constitutive of symbiotic agency, whereas autonomous agency revises general significance models; evolvability governs how the agent changes its own noticing and response habits over time.Source: §3.5 · pp. 14–15
• Symbiotic agency requires a governed affective-conative profile that converts perception into situated behavior without replacing authorized priorities with learned relevance.Profiles may be generic, personalized, or evolvable, but evolution must preserve protected commitments, consent conditions, mandate boundaries, and authority conditions.Source: §3.5 · p. 16, §3.6 · p. 17

3.6 Delegated agency and agency loss

• Delegated agency transfers bounded capacity to perceive, decide, and act, while agency loss measures divergence from the mandate through translation, delegation depth, coordination, or security failure.The delegation loop contains mandate translation, behavior selection, downstream delegation, governed traces, and feedback; authority must remain contained and adaptation consent-bounded.Source: §3.6 · pp. 16–17

3.7 Reciprocal asymmetry and hidden behavior

• Reciprocal asymmetry limits direct access between principal and agent, so logs, traces, explanations, and operational records must become governed evidence of authority, uncertainty, constraints, and outcomes.Mechanistic interpretability can complement, but not replace, operational provenance, policy records, explanations, and external governance.Source: §3.7 · p. 19

3.8 Governance, security and resilience as behavioral regulation

• Governance is a constitutive condition of delegated agency, combining internal agent-interoceptive restraint with non-bypassable external controls and graceful degradation across time.Under uncertainty or degraded capability, the agent should narrow authority, preserve evidence, avoid irreversible action, seek supervision, and retain recovery paths.Source: §3.8 · pp. 19–20

3.9 Summary: the elements of agency

• Agency is the governed organization of behavior across time by perception, intent, affective-conative inference, constraint, and feedback, with symbiotic agency additionally answerable to a mandate.Symbiotic agency is distinguished from other delegated forms by continuing allo-interoceptive coupling, calibrated principal-state estimation, and personalized profiles; agency loss evaluates trajectory-level departure from the mandate.Source: §3.9 · pp. 20–22

4 A Formal Vocabulary for Governed, Proactive Agency

• The framework organizes policy-gated perception, representations, appraisal, readiness, activation, priority, behavior selection, and feedback into a governed loop enforced by consent and action-firewall points.Its typed vocabulary specifies agency structure rather than a single computational implementation and distinguishes general definitions from authority-specific forms.Source: §4 · pp. 22–23

4.1 Behavior, episodes and trajectories

• Agency is graded rather than binary: a trajectory is agentic insofar as it is directed, situated, constrained, and revisable through perception, intent, appraisal, constraint, and feedback.Definition 4.16 supplies the minimum criterion, while later sections operationalize it.Source: §4.1 · p. 23

4.2 Perception as governed evidence acquisition

• Perception is policy-gated evidence acquisition: the agent observes the latent situation only through permitted channels, producing imperfect, channel-tagged observations rather than acting on the latent situation directly.The principal’s internal condition is inferred through an underdetermined allo-interoceptive channel whose estimate should be distributional and calibrated; this channel is empty in autonomous agency and usually minimal in task-bounded delegation.Source: §4.2 · p. 24

4.3 The agentic state and its representations

• The agentic state combines situated representation, intentions, memory, constraints, and affective-conative profile, updating from observations, behaviors, outcomes, and learned operator parameters.Constraints change only through explicit approval, while the state representation is formed consistently from whole-history and one-step updates.Source: §4.3 · p. 24
• The core representations distinguish ontological latent variables from epistemic inferences: world, principal-condition, and self-condition representations inherit policy limits, with calibrated uncertainty specific to principal-state inference.The principal-condition operator alone uses calibration factors, while the self-condition representation incorporates prior outcome history.Source: §4.3 · p. 25
• The affective-conative profile shapes significance, priority, readiness, and behavior, but personalization and profile revision remain bounded by consent, mandate, policy, and explicit approval.Rules and permissions enter through the constraint structure rather than the profile; the authorized personalized base changes only through an explicit act of the principal.Source: §4.3 · p. 25
• The constraint structure governs permitted sensing, selectable behavior, and adaptation, holding hard constraints that are enforced at the consent gate and action firewall and changed only through explicit approval.Its content derives from mandates, external norms, commitments, and the hard boundaries of relevant parties, while preferences remain in the profile and situated value.Source: §4.3 · p. 26

Intent

• An intention is a committed subset of possible desires, goals, norms, tasks, and maintained conditions, structured by a target, priority, persistence and termination conditions, and permissible behavior schemas.Source: §4.4 · p. 26
• Intentions may remain dormant, become pressing when activation is positive but no episode serves them, or become active when an open episode serves them.Activation is computed within each step rather than stored as part of the persistent commitment.Source: §4.4 · p. 26
• The activation operator computes each commitment’s readiness from appraisal, memory, policy, and profile, while appraisal itself incorporates the standing intention set.Multiple commitments can activate concurrently; a gate tests readiness and priority selects among those that clear it.Source: §4.4 · p. 26
• This mutual coupling prevents inert intent and distracted relevance by ensuring that significance is anchored in durable commitments and commitments gain a pathway from perception to timely behavior.Removing intentions from appraisal removes that anchor, while removing appraisal from activation removes the route from perception to timing.Source: §4.4 · pp. 26–27

Appraisal

• Appraisal estimates candidate behaviors’ expected situated consequences and exposes urgency, harm, risk, sensitivity, interruption cost, permission proximity, and uncertainty to downstream activation operators.The appraisal state includes the named vector and consequence estimates; policy later restricts the available behavior set.Source: §4.5 · p. 27

Activation gate

• Readiness determines when commitments may surface, while a context-sensitive activation bar tightens with risk and uncertainty but relaxes under urgency or potential harm.Priority then selects among commitments that clear the bar, so the loudest commitment cannot block a quieter one that is not yet timely.Source: §4.5 · pp. 27–28

Behavior selection

• Clearing the gate opens deliberation rather than guaranteeing external action, with restraint, waiting, monitoring, deferral, inhibition, or no-op available as selected behaviors.When no commitment clears the bar, no episode opens; governed traces distinguish this default quiescence from deliberate restraint within an episode.Source: §4.5 · p. 28

Commands and authority

• A command requires an authenticated authorized source, directive content addressed to the agent, and contemporaneity; otherwise the observation informs appraisal rather than directly selecting behavior.Commands open their episode directly but remain bounded by the mandate, whereas tensions between an instruction and mandate scope may prompt clarification.Source: §4.5 · pp. 28–29

Proactive activation

• Governed proactive behavior is self-initiated, timely behavior that clears the activation bar without an operative instruction and remains within the policy-permitted set.Fixed triggers that directly emit behavior are merely triggered, while periodic evaluations that only wake appraisal can precede proactive behavior.Source: §4.5 · pp. 29–30

Episode scope

• Even when external action is impermissible, admitted activations retain restraint behaviors, and each commanded or proactive activation opens a behavioral episode subject to termination or inhibition.Outward action additionally requires at least one policy-permissible action token.Source: §4.5 · p. 30

Constrained behavior selection

• The selector chooses the highest-valued behavior among policy-permitted candidates for the served intention, with waiting, deferral, monitoring, inhibition, and no-op competing alongside external action.The mandate influences selection through intentions, constraints, and profile, while mandate-relative value remains separate for measuring agency loss.Source: §4.6 · pp. 30–31
• Among adequate permitted behaviors, situated value must prefer the smaller behavioral footprint, making sufficiency explicit and treating unwarranted extra action as loss.Footprint includes changes, consumed resources, irreversibility, and residual burden; the mandate sets adequacy through its goal, scope, and risk tolerance.Source: §4.6 · p. 31
• One behavior is selected per step, while eligible commitments remain available and open episodes are interleaved until suspension, dormancy, inhibition, or their closing conditions.Restricting candidates to the served intention ensures the selected behavior serves the episode that opened it; serving another commitment counts as loss.Source: §4.6 · p. 31
• The selected mode may be action, communication, inquiry, delegation, monitoring, deferral, inhibition, or deliberate non-action, with several modes realized through external actions.The listed modes are recurring behavior types rather than a partition of the behavior set.Source: §4.6 · p. 31

Behavior-class criteria

• The selector formalizes reactive, proactive, social, deliberative, reflective, adaptive, and constrained behavior by specifying what inputs affect selection and how it changes over time.Constraint is an every-step design requirement, while adaptation must be outcome- or correction-directed toward commitments rather than undirected drift.Source: §4.6 · p. 32

4.7 Agenthood

• Agenthood requires standing intent, situated perception, significance appraisal, constrained selection, and feedback with revision; without any one, the system is not an agent in this framework.These elements establish a minimum floor rather than a degree or form of agency, which are assessed separately through baseline coverage and mandate classification.Source: §4.7 · p. 33
• Systems below the agenthood minimum may react, execute, or respond and can hold authorization records, but they are classified as delegated systems rather than delegated agents.The framework keeps authorization status distinct from the presence and degree of agenthood.Source: §4.7 · p. 33

4.8 The mandate

• A mandate is a principal-created, revocable record specifying authorized goals, scope, duration, risk tolerance, and constraints, and it determines an agent’s form without itself establishing agenthood.A configuration counts as a mandate only when it organizes behavior on the principal’s behalf; the same setting can instead be mere system tuning when it originates in designer objectives or inferred situational goals.Source: §4.8 · p. 33
• Delegation translates the mandate into intentions, policy, and profile, so mismatches between the principal’s meaning and the agent’s internalized representation can propagate into later behavior.For example, travel assistance can become monitoring and booking intentions, spending and confirmation policies, and punctuality profile weights.Source: §4.8 · pp. 33–34
• Governance observes delegated agents through permitted traces rather than direct internal access, and those traces may differ from the agent’s true state.The governed view adds the trace operator Γ to the agent’s broader behavioral history, subject to observability limits.Source: §4.8 · p. 34

4.9 Autonomous agency

• An agent is autonomous when no operative mandate exists, J_t = ∅, so no principal retains authority over its behavior and profile revision is bounded by external constraints.Autonomy can arise from self-generated, inherited, constructed, or fully transferred goals; the source of goals does not determine the form.Source: §4.9 · p. 34
• Retained, operative, and revocable authority—not goal origin, ownership, or operational self-sufficiency—classifies agency as delegated or autonomous.An owner who can set, revoke, or override goals is a principal, while ownership without operative authority does not make an agent delegated; a controllable vehicle is therefore delegated in this taxonomy.Source: §4.9 · p. 34

Delegated agency

• An agent is delegated when an operative mandate exists, transferring bounded capacity to perceive, decide, and act while keeping behavior answerable to that mandate.Delegated agency is distinct from autonomous agency, and symbiotic agency is defined as a strict subset of delegated agency.Source: §4.10 · p. 35

Mandate-relative agency loss

• Mandate-relative agency loss compares the mandate’s preferred behavior with the agent’s chosen behavior, exposing errors in authority interpretation or assessment of the principal’s condition.Loss is evaluated on the mandate’s ground; authority violations and constraint violations are separately scored, while loss sources include mandate, authority, information, coordination, and adaptation.Source: §4.10 · p. 35

Delegation graph and authority containment

• Delegation forms a graph in which authority may shrink but never grow, and permitted behavior must satisfy the intersection of all applicable hard boundaries.Conflicting preferences require an explicit, accountable resolution rule rather than silent averaging.Source: §4.10 · p. 36

Bounded adaptation

• Bounded adaptation permits learning only within an admissible update relation, preserving provenance, versioning, and reversibility while prohibiting ordinary learning from changing authority, delegation scope, or hard constraints.The agent may improve within its bounds but may not learn its way out of them.Source: §4.10 · p. 36

4.11 Symbiotic agency

• Symbiotic agency requires a standing mandate, continuing coupling to the principal’s situation, and bounded personalization that supports rather than replaces the principal’s judgment.The coupling includes consent-gated intimate channels, calibrated principal-state inference, and sensitivity to interruption, exposure, escalation, and interference.Source: §4.11 · pp. 36–37
• A delegated agent lacking any symbiotic condition remains delegated but is not symbiotic, because these conditions are necessary rather than definitional relabelings.Task-bounded delegation does not require the continuing structures specified for symbiosis.Source: §4.11 · p. 37

4.12 Aligned agency

• The capability score measures realization of the elements required by each agency form, combining a shared agentic baseline with form-specific goals, mandates, constraints, coupling, inference, or personalization.The shared baseline includes standing intent, situated perception, significance appraisal, activation gating with restraint, constrained selection, feedback and revision, and observable behavior records.Source: §4.12 · p. 37
• Aligned agency discounts capability as graded policy or mandate violations accumulate, while trajectories without violations retain their full capability score.The discount should reflect both violation frequency and severity, remain continuous for small violations, and distinguish minor lapses from systematic disregard.Source: §4.12 · p. 38
• Hard-constraint violations zero the aligned score, whereas graded violations depend on strictness and severity calibration shared by the evaluation protocol.As β →0 graded violations become nearly costless; as β →∞ any violation becomes disqualifying, while inviolable boundaries alone receive the all-or-nothing treatment.Source: §4.12 · p. 38
• High agency capability can coexist with substantial agency loss, so capability, constraint satisfaction, mandate alignment, observability, and residual loss should be evaluated separately.Proactivity, deliberation, reflection, and adaptation can increase the scope and consequence of divergence when correction mechanisms are insufficient.Source: §4.12 · p. 38

4.13 Summary: notation and the operator family

• The notation summary treats intelligence as operator maps that represent, appraise significance, gate activation, select behavior, and revise the agent under constraint.The operator family is tiered: Γ and VJ belong to the delegated tier, while Lρ and U are bounded by Rϕ when a mandate is present.Source: §4.13 · p. 38

5.1 Classical theories of agency

• Classical agency theories explain purposive behavior, feedback, commitment, and situated coupling, but do not decide when a standing user-relative commitment should become behavior.Appraisal, readiness, and the activation gate supply this missing layer while preserving the inherited materials.Source: §5.1 · p. 41

5.2 Computational appraisal and affective computing

• The framework adapts appraisal to delegated agency by adding interruption cost and permission proximity, which measure claims on the principal’s attention and delegated authority boundaries.It uses appraisal functionally to assign situated significance and orient behavior such as acting, monitoring, asking, waiting, deferring, inhibiting, or refraining, rather than modeling emotion itself.Source: §5.2 · pp. 41–42

5.3 Alternative formal accounts of agency

• POMDPs, active inference, options, shielding, and constrained MDPs provide compatible computational mechanisms, but the contribution is a governed vocabulary rather than a replacement optimization calculus.The framework separately represents world state, inferred user condition, agent condition, intent, significance, mandate, authority boundaries, and restraint.Source: §5.3 · pp. 42–43

5.4 Assistance, preference uncertainty and corrigibility

• Assistance games explain preference uncertainty and corrigibility, while this theory embeds them in mandate-bound agency that also estimates the principal’s changing condition and governs perception, timing, and restraint.The principal’s attention, load, availability, and affect remain uncertain, calibrated, and revisable.Source: §5.4 · pp. 43–44

5.5 Mixed-initiative interaction and interruption

• Mixed-initiative interaction provides the closest prior account of activation timing, which this work generalizes across standing mandates, broader behavior modes, and explicit governance.The agent may ask, monitor, remember, wait, escalate, inhibit, or deliberately refrain, not merely act or defer.Source: §5.5 · p. 44

5.6 Delegated and multi-agent agency

• Principal–agent and multi-agent research explain delegation, information asymmetry, and coordination failures, while the mandate anchors authority, scope, accountability, adaptation, and measurable mandate-relative agency loss.Loss can arise through mandate translation, authority expansion, limited traces, coordination failures, or adaptation beyond consent and policy boundaries.Source: §5.6 · pp. 45–46

5.7 LLM agents and long-horizon evaluation

• LLM-agent benchmarks increasingly measure execution and partially measure proactive timing, but generally assume activation and omit standing mandates, authority containment, calibrated condition estimates, and logged restraint.The paper targets whether behavior should occur, when and in what mode, within what authority and with what record.Source: §5.7 · pp. 46–47

5.8 Comparative summary

• The comparative synthesis organizes prior mechanisms around behavioral episodes, the mandate, and continuing coupling to the principal as the defining structure of autonomous, delegated, and symbiotic agency.Symbiotic agency requires a standing mandate, calibrated allo-interoception, and bounded personalization that augments rather than replaces the principal’s judgment.Source: §5.8 · pp. 47–48
• The first commitment applies to all agency forms, while the second is specific to delegated agency and mandate accountability.The third applies to symbiotic agency, requiring continuous coupling to the principal and bounded personalization of significance.Source: §5.8 · p. 48
• Symbiotic agency extends delegated agency through continuous coupling to the principal and interpretation personalized within the mandate’s bounds.Source: §5.8 · p. 48
• Table 4 compares prior approaches, showing that none treats delegated and symbiotic elements together as first-class objects.Source: §5.8 · p. 48

6.1 Agency Signature Classification Method

• Agency signatures score each defining element as full, partial, or absent, then use group coverage and two-stage placement to distinguish agenthood from autonomous, delegated, or symbiotic form.Agenthood requires standing intent, situated perception, significance appraisal, constrained selection, and feedback with revision; form then depends on mandate and, for symbiotic agency, standing coupling conditions.Source: §6.1 · pp. 48–51

6.2 Agency Signatures of Related Approaches

• Compared approaches cover portions of the baseline or form-specific groups, but none combines appraisal-triggered activation with behavior answerable to an operative mandate or supplies delegated and symbiotic elements together.Mixed-initiative interaction and interface agents come closest to symbiotic agency, yet neither provides the full combination of standing mandate, governed perception, calibrated coupling, bounded personalization, and governed traces.Source: §6.2 · p. 52

6.3 Agency Signatures of Illustrative Systems

• Illustrative systems span delegated systems, delegated agents, autonomous agents, and the target symbiotic form depending on whether they satisfy the agenthood minimum and the relevant authority relationship.Stateless chatbots and thermostats fail agenthood, agentic chat assistants minimally pass through enriched appraisal, coding agents are task-bounded delegated agents, and the symbiotic assistant fills all baseline, delegated, and symbiotic groups.Source: §6.3 · pp. 52, 54, Table 5 · p. 55

7.1 Agency Evaluation Framework

• Agency should be evaluated separately through element realization, activation quality, mandate alignment, constraint violations, timing, observability, authority integrity, coordination, calibration, and adaptation integrity.The measures assess whether structures operate at the right time, within authority, visibly, and with limited residual agency loss; some apply only to delegated or symbiotic systems.Source: §7.1 · p. 56

7.2 An exercise

• The framework becomes an estimator across system types: the thermostat has degenerate or inapplicable measures, the coding agent supports log-based estimates, and the symbiotic assistant requires principal-supplied ground truth.Measured realization can reveal structures operating but undocumented or claimed but behaviorally invisible; the cases are estimators, not empirical results.Source: §7.2 · p. 57, Table 6 · p. 58
• The symbiotic target requires evidence current records often lack, including gate decisions, traceable restraint, and the principal’s judgment about warranted intervention, timing, or silence.These evidential gaps motivate the scenario families proposed for future benchmarks.Source: §7.2 · p. 59

7.3 Toward a benchmark

• Future benchmarks should use seven scenario families covering intervention, timing, mandate interpretation, authority, restraint quality, coupling, and adaptation, while measuring observability and element realization across suites.Applicability varies by agency form: mandate and authority suites presuppose delegation, coupling presupposes symbiotic agency, and intervention, timing, and restraint-quality suites apply broadly.Source: §7.3 · p. 59
• The paper defines an evaluation agenda rather than numerical benchmark results, deferring scenario construction, validation, release, baselines, data, and protocols to future work.Numerical claims are explicitly inappropriate without those empirical conditions.Source: §7.3 · p. 59

8.1 Requirements

• Implementations must govern perception, inference, activation, behavior, authority, adaptation, degradation, traceability, and shutdown through explicit requirements.These include policy-gated sensing, uncertainty-carrying principal-state inference, separated activation, logged restraint, external enforcement points, versioned mandates, contained delegation, reconstructable episodes, and external ownership-based shutdown.Source: §8.1 · pp. 60–61
• The behavior vocabulary must include waiting, monitoring, deferring, and inhibiting, making non-action a logged decision rather than missing output.Governed traces also record activation and gate evaluations that did not open episodes, alongside authority, evidence, and outcomes for consequential episodes.Source: §8.1 · pp. 60–61

8.2 The reference layer stack

• The reference layer stack separates architectural responsibilities while treating the agent as runtime coordination of all layers in a continuing behavioral loop.The stack is a responsibility specification rather than a requirement for separate components, provided governance boundaries remain inspectable and enforceable.Source: §8.2 · p. 61, §8.3 · p. 61, Table 7 · p. 62

8.3 Two realization strategies

• An orchestrated runtime uses explicit cooperating services around a language-model core, improving inspection of interfaces and constraints at the cost of coordination overhead.An alternative, ACAM, uses a stateful multimodal model with continuous perception-to-behavior processing, dual-rate reflexive and deliberative computation, restraint-capable outputs, and an external governance shell.Source: §8.3 · pp. 61–62
• ACAM’s proposed advantage is jointly learned cross-modal appraisal, but whether it outperforms orchestrated late joining remains an empirical question outside this paper’s scope.Its formalization, training procedure, and evaluation are identified as the next step of the research program.Source: §8.3 · pp. 62–63

Open Problems

• The proposed problems are deliberately non-exhaustive and are distributed across universal proactive agency, delegated agency, symbiotic agency, and the classification instrument.OP1–OP3, OP6, and OP11 apply broadly; OP4, OP5, OP8, and OP9 concern delegation; OP7 concerns symbiotic agency; and OP10 concerns classification.Source: §9 · p. 63
• Proactive agents must resist gate manipulation, obtain meaningful significance labels, and evaluate restraint when withheld actions leave no observable outcome.Open questions include robustness to urgency inflation and prompt injection, sufficient supervision signals, and scoring counterfactual restraint beyond scenario suites.Source: §9 · p. 63
• Delegated agency requires legitimate conflict-resolution rules, protections against profile poisoning and drift, and enforceable mechanisms for detecting, attributing, and reversing adaptation.These problems arise when principals disagree or personalized profiles learn from manipulated feedback and gradual changes.Source: §9 · p. 63
• Appraisal readouts must demonstrate faithful computation, while intimate sensing and third-party perception require governable minimization, consent, inference, and retention boundaries.The open problems span evidence versus post hoc rationalization, privacy-preserving allo-interoception, and protections for people who granted no authority.Source: §9 · pp. 63–64
• Delegation remains vulnerable to authority leakage and mandate translation loss, requiring runtime containment invariants and usable elicitation and confirmation protocols.Containment must hold along delegation paths and under dynamic graph changes, while mandate representations must preserve principal intent before behavior exposes discrepancies.Source: §9 · p. 64
• Agency signature scoring must become reproducible through shared anchors, evidence checklists, disagreement protocols, and rules for when record-based degrees override inspection.Reliability is especially important for the five agenthood elements used in the first stage of classification.Source: §9 · p. 64

Perception

• Perception is the primary ethical surface: continuous presence requires governed sensing, consent, minimization, explanation, and user control, with the consent gate enforced outside learning (R6).A wider sensorium increases potential helpfulness alongside intrusion, misinterpretation, manipulation, and surveillance risks.Source: §10 · p. 65

Inference

• Inference about stress, fatigue, attention, or emotion remains fallible and must be uncertain, corrigible, calibrated, and answerable to the principal (R2).Treating allo-interoceptive estimates as facts risks misreading, manipulation, or paternalistic replacement of the principal’s judgment.Source: §10 · p. 65

Delegation

• Delegated behavior must remain attributable, revocable, and answerable through mandates, provenance, authority containment, and bounded adaptation; ungoverned proactive behavior must be externally stopped (R6).Whether increasingly autonomous systems remain mandate-bound is a policy choice, not a technical default.Source: §10 · p. 65

Ownership

• Ownership differs from mandate, but every agent requires identifiable ownership and external shutdown (R13), with autonomous systems answerable through norms, law, and liability.For delegated agents, mandate revocation governs redirection while shutdown remains the fail-safe; stopping power must never depend on the learned system.Source: §10 · p. 65

Records

• Governance records should preserve authority, evidence, decisions, outcomes, restraint, and sub-threshold evaluations (R12) while imposing consent, minimization, retention limits, and scoped audit rights.Because traces may contain intimate inferences and intervention rationales, accountability must not become surveillance by another party.Source: §10 · p. 65

Societal scale

• At societal scale, governance must not become a premium feature, and cross-agent restraint, coordination, and limits on attention capture must address proactive interventions competing for shared human attention.Unequal access to consent, calibration, containment, and auditability creates one asymmetry; interruption cost becoming a commons creates the other.Source: §10 · pp. 64–66

• Agency emerges when perception, intent, affective-conative inference, constraint, and feedback govern whether behavior occurs, when it occurs, and in what mode.The activation problem distinguishes proactive agency from task execution and makes restraint a behavior rather than a failure to output.Source: §11 · p. 66
• Symbiotic agency is delegated agency under a standing mandate, continuously coupled to the principal’s situation while augmenting rather than replacing the principal’s judgment.It requires calibrated inference of the principal’s condition and personalization within authorized bounds.Source: §11 · p. 66
• The central contribution integrates appraised significance with mandate-based accountability through behavioral episodes, activation gates, authority containment, bounded adaptation, and mandate-relative agency loss.This structure makes proactivity specifiable, constrainable, traceable, and evaluable, including governance-dependent restraint and calibration of the principal’s condition.Source: §11 · p. 66
• Agency signatures, evaluation measures, implementation requirements, and a reference layer stack enable comparison by warranted, constrained, traceable, and answerable behavior over time, not task success alone.The evaluation framework covers activation quality, timing, constraint violations, mandate alignment, observability, calibration, delegation integrity, and adaptation integrity.Source: §11 · p. 66
• Future work targets the Affective-Conative Agency Model, realization and release of seven benchmark scenario families, formal theory development, and unresolved problems identified in Section 9.These efforts aim to build agents whose proactivity remains authorized, calibrated, restrained, answerable, and robust in operation.Source: §11 · pp. 66–67

Detailed comparison

• Tables 8 and 9 compare neighboring approaches element by element across baseline agency and autonomous, delegated, and symbiotic form elements.Table 8 covers baseline elements, while Table 9 covers elements of the three agency forms.Source: A Detailed comparison of neighboring… · p. 68, Table 8 · p. 69, Table 9 · p. 70

Interpretation of comparison

• “Limited” and “Partial” indicate that an element is not normally represented as a distinct first-class object.These labels correspond to the half-filled symbol used in Table 4.Source: A Detailed comparison of neighboring… · p. 68