PaperlayerView exact revision on arXiv

Source-linked AI summary

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

João Dias Ferreira

arXiv:2609.07741v1cs.AI

TL;DR

Persistent assistants need to decide whether, when, and how to act rather than merely execute prompted tasks. This paper develops a governed-agency framework organized around perception, intent, appraisal, constraint, and feedback, and proposes concepts and instruments for assessing authorized, accountable assistance.

  • Problem

    Current agent systems can execute sophisticated delegated tasks but lack an integrated account of activation relative to ongoing context, standing commitments, mandates, and user condition.

  • Method

    The paper develops a conceptual and formal framework of autonomous, delegated, and symbiotic agency, centered on governed behavioral episodes and supported by classification, evaluation, benchmark, and architecture instruments.

  • Results

    The framework integrates appraised activation with operative mandates, authorized perception, authority containment, traceable restraint, and constrained adaptation.

  • Takeaways & Limitations

    The account provides a basis for specifying and assessing assistance as warranted, timely, authorized, and answerable beyond task completion.

  • Takeaways & Limitations

    Inferring the principal’s internal condition is limited by the agent’s indirect, noisy, incomplete, delayed, or policy-restricted evidence.

Abstract

from arXiv · show

Persistent AI assistants are intended to extend human attention, memory, and coordination across changing digital and physical environments. To be truly useful they must do more than just act when asked. They must decide on their own whether a situation warrants behavior at all, when it does and in what mode, whether to act, ask, monitor, defer or deliberately refrain. We call this the activation problem. Research on commitment, appraisal, mixed-initiative interaction and delegation each illuminates part of it, but none ties situated activation to continuing authorization and accountability. This paper develops a conceptual and formal framework for governed proactive agency, organizing behavior across time through perception, intent, affective-conative appraisal, constraint, and feedback. It distinguishes autonomous and delegated agency and defines symbiotic agency as delegation under a standing, revocable mandate, with continuing coupling to the principal's situation, calibrated inference of their condition, and bounded personalization. The distinctive contribution is an integrated account linking activation decisions to authorized perception, behavior selection, authority containment, traceable restraint, and constrained adaptation, with behavioral episodes as the unit of analysis. Through an agency classification method, an evaluation framework, proposed benchmark scenarios, and a reference architecture, the account provides a basis for specifying and assessing whether assistance is warranted, timely, authorized, and answerable beyond task completion alone. It is intended to guide the development and evaluation of always-present personal assistants and embodied support systems that augment human capabilities while preserving the principal's authority and judgment.

1 Introduction

The paper frames agency as organized behavior across time and identifies the Activation Problem: deciding whether, when, and how to behave, not merely executing assigned tasks. It distinguishes autonomous, delegated, and symbiotic agency, then proposes governed structures for authorized, accountable assistance.

  • The Activation Problem: The Activation Problem is deciding whether behavior is warranted, when it should occur, and whether to act, ask, monitor, defer, escalate, or refrain.Existing agentic capabilities mainly determine what to do and how once a task has been given.
  • Agency as organized behavior: Agency is an organizational property in which perception, intent, affective-conative inference, constraint, and feedback direct behavior across time.These elements organize behavior so it is directed, situated, constrained, and revisable.
  • Agency as organized behavior: The agenthood minimum requires all five elements at least in part, while fuller agency additionally includes activation choice, restraint, and an observable record of conduct.Agency is treated as a matter of degree between the minimum and complete presence of these capacities.
  • Forms of agency: Symbiotic agency is delegated agency under continuing coupling to a principal, where the agent infers significance and may act, wait, ask, escalate, or refrain to augment judgment.The paper contrasts this with autonomous agents pursuing their own goals and bounded delegated agents completing transferred tasks.
  • Governance and accountability: A mandate makes symbiotic coupling answerable, while ownership separately concerns who operates and can stop the agent.Without an explicit mandate, continuous presence can drift toward unbounded autonomy, intrusion, or surveillance.
  • Scope and contributions: The paper contributes a taxonomy, formal vocabulary, classification method, evaluation framework, benchmark direction, and reference layer stack for governed agency.Its stated aim is a conceptual and formal account rather than a single implementation or empirical result.

2 A Genealogy of Agency as Organized Behavior

The genealogy moves from stimulus–response explanations toward agency as purposive, situated, committed, and socially organized behavior. The paper argues that long-horizon task execution still leaves open how episodes are activated, constrained, revised, and suspended over time.

  • From reactivity to purposiveness: Early stimulus–response accounts explained local reaction but struggled with behavior organized around aims when environments vary or restraint is appropriate.They clarified reactivity more than directed, situated, constrained, and revisable agency.
  • From reactivity to purposiveness: Tolman shifted analysis toward purposive acts and argued that goal-directed behavior is a property of the whole organism, supported by internal cognitive organization.This challenged reducing agency to isolated movements or local stimulus–response links.
  • Feedback and computation: Cybernetics made purposiveness mechanizable by using feedback to reduce the gap between current and target states, but it presupposed an active target.The paper uses this boundary to motivate the question of how targets or episodes become active.
  • Commitment and situatedness: BDI theory contributed durable intentions that constrain future deliberation, yet a standing intention may remain dormant, activate, be deferred, inhibited, or revised.Thus commitment alone does not determine when intention should become behavior.
  • Commitment and situatedness: Situated approaches emphasize coupling to changing conditions, but continuous reaction to salient signals still lacks prioritization, inhibition, mandate sensitivity, and significance.The paper therefore treats situatedness as necessary but insufficient for stronger agency.
  • Contemporary agents and behavioral episodes: Most LLM agents remain organized around initiating events and can execute long-horizon tasks, while activation relative to standing commitments and mandates remains open.The paper identifies the behavioral episode—not the isolated action—as the unit for examining noticing, evaluation, selection, inhibition, revision, and restraint.

3 The Theory: Agency as Organized Behavior

Agency is organized behavior across time, not a collection of capabilities or isolated actions. The theory links persistent commitments, affective-conative appraisal, authorized personalization, governance, and feedback to behavior that may include intervention, restraint, and revision.

  • Behavior, action and restraint: The agentic unit is the organized behavioral episode, encompassing action, timing, sequencing, inhibition, correction, escalation, waiting, and deliberate non-action.An isolated action may be technically impressive yet weakly agentic when detached from aims, constraints, and revision.
  • Agency as organized behavior: Agency becomes more than capability when persistent commitments organize action selection across time and update in response to feedback.The paper therefore treats agency as an organizational property rather than the sum of models, memory, tools, planning, or skills.
  • Affective-conative inference: Affective-conative inference evaluates context against commitments, risks, permissions, priorities, preferences, and possible outcomes before intention becomes active behavior.Conation names the action-oriented movement from perception and valuation toward intention, choice, and behavior.
  • Affective-conative inference: Proactivity includes acting, asking, remembering, monitoring, delaying, lowering priority, or deliberately refraining when a situation warrants those modes.Restraint is treated as a possible output of governed appraisal rather than as an absence of agency.
  • Symbiotic agency: Symbiotic agency uses personalized affective-conative inference under a mandate to connect continuous coupling with answerability to the principal.Allo-interoceptive evidence supplies coupling, while the mandate governs what that evidence means for behavior.
  • Governance and boundaries: Personalization and delegation remain bounded because intimate sensing increases risks of intrusion, misinterpretation, manipulation, surveillance, and unauthorized priority substitution.The sensorium must be governed through consent, locality, policy, explainability, minimization, user control, and distinguishable mandate boundaries.
  • Delegated agency: Delegated agency transfers a limited capacity to perceive, decide, and act, while multi-agent delegation can compress context, misinterpret permissions, and complicate responsibility.The delegation loop translates mandates into intent, policy, and profile, selects behavior, supports downstream delegation, and returns governed traces and outcomes.
  • Governance and boundaries: Governance preserves delegated behavior as reliable, bounded, observable, and recoverable by recording mandates, constraining authority, detecting divergence, and enabling correction or revocation.Digital agents can retain records of tool calls, memory updates, policy checks, delegated tasks, messages, state changes, explanations, and outcomes.

4 A Formal Vocabulary for Governed, Proactive Agency

The formal vocabulary treats agency as a governed loop whose distinct representations, perceptions, constraints, and adaptive state organize behavior across time. It remains implementation-agnostic while separating general agency structures from delegation-specific authority and governance.

  • 4 A Formal Vocabulary for Governed, Proactive Agency: Figure 3 organizes policy-gated perception, representations, appraisal, readiness, activation, priority, behavior selection, and feedback into a governed loop.The vocabulary specifies distinct elements without committing to one computational substrate.
  • 4 A Formal Vocabulary for Governed, Proactive Agency: The vocabulary is tiered: general definitions describe autonomous and delegated agents, while mandate-based definitions add delegation-specific governance structures.The general tier leaves room for delegation without presupposing it.
  • 4.1 Behavior, episodes and trajectories: Behavior includes external action, communication, tool use, inquiry, internal updates, waiting, inhibition, and deliberate restraint, while episodes are contiguous trajectory segments.Agentic trajectories are directed, situated, constrained, and revisable; agency above the minimum is graded.
  • 4.2 Perception as governed evidence acquisition: The latent situation combines the external world, the principal’s latent condition, and the agent’s internal condition, which the sensorium accesses through distinct channels.Allo-interoceptive inference is underdetermined and should remain distributional with calibrated uncertainty.
  • 4.2 Perception as governed evidence acquisition: Policy-gated perception maps the latent situation, preceding behavior and outcome, and noise into observations using only permitted sensing channels.The perception operator explicitly incorporates noise, incompleteness, latency, and distortion.
  • 4.3 The agentic state and its representations: The agentic state contains situated representations, intentions, memory, constraints, and an affective-conative profile, updated from observations, behavior, outcomes, and learned parameters.Constraints may change only through explicit approval when a mandate is present.
  • 4.3 The agentic state and its representations: Core representations distinguish ontological latent variables from epistemic estimates of the world, principal, and agent, formed by operators with defined evidence inputs.The agent’s profile includes generic, personalized, and evolvable components, while rules and permissions remain in the constraint structure.

4.4 Intent

Intentions provide durable commitments, while affective-conative appraisal determines which commitments become behaviorally relevant under present conditions. Their mutual coupling prevents commitments from becoming inert or situational relevance from becoming unanchored.

  • 4.4 Intent: An intention is a committed target or maintained condition with priority, persistence and termination conditions, and permissible behavior schemas.Intentions may remain dormant while continuing as standing commitments.
  • 4.4 Intent: The activation operator uses representations, appraisal, memory, policy, and profile to determine how strongly a standing commitment presses to become behavior now.Appraisal also incorporates the standing intention set, creating mutual coupling between commitments and significance.
  • 4.4 Intent: Intent supplies stable behavioral organization, whereas affective-conative inference determines whether a commitment should become active under current conditions.This distinction separates durable commitment from situated relevance.
  • 4.4 Intent: Mutual reference between intention and appraisal avoids inert intent and distracted relevance by linking significance to commitments and commitments to present conditions.Removing either side would permit commitments without behavioral influence or significance without durable anchoring.

4.5 Appraisal, priority, readiness and the activation gate

The activation gate turns appraised significance into timely, authorized behavior by combining readiness, context-sensitive thresholds, priority, and command verification. It also treats waiting, monitoring, deferral, inhibition, and no-op behavior as deliberate outcomes when action is not warranted.

  • 4.5 Appraisal, priority, readiness and the activation gate: Appraisal evaluates the current situation’s significance for behavior using representations, intentions, memory, policy, profile, and expected situated consequences.The appraisal vector summarizes named significance components and candidate-behavior consequences under uncertainty.
  • 4.5 Appraisal, priority, readiness and the activation gate: Readiness measures each pressing commitment’s disposition to surface, and only commitments exceeding the context-sensitive activation bar become eligible.The threshold depends on appraisal, estimated principal and agent conditions, policy, and profile.
  • 4.5 Appraisal, priority, readiness and the activation gate: Priority selects among eligible commitments by comparing activation levels weighted by their priorities, while readiness governs timing rather than choice.A commitment that is more urgent but not yet timely does not block a quieter eligible one.
  • 4.5 Appraisal, priority, readiness and the activation gate: A cleared gate can open an episode whose selected behavior is action, restraint, waiting, monitoring, deferral, inhibition, or no-op; sub-threshold evidence instead updates standing operation.Deliberate non-action is recorded within an episode when selected, whereas no cleared commitment opens no episode.
  • 4.5 Appraisal, priority, readiness and the activation gate: A command requires an authenticated authorized source, directive content, and contemporaneity; otherwise the observation enters appraisal as evidence rather than instruction.Imperative-looking unauthorized content, including phishing messages, is not allowed to determine activation as a command.
  • 4.5 Appraisal, priority, readiness and the activation gate: Proactive behavior is self-initiated and timely, while governed proactivity additionally requires policy permission and therefore excludes unconstrained activation.The activation bar may be vulnerable to adversarial urgency inflation, identified as an open problem.

4.6 Behavior selection under constraint

Behavior selection chooses one permitted behavior for the winning commitment by optimizing expected situated value under active constraints. The framework makes restraint and sufficiency explicit, treating behavior beyond the mandate’s warranted bar as a loss rather than diligence.

  • 4.6 Behavior selection under constraint: Policy defines the permissible behavior set, and selection serves the intention that opened the episode using only behaviors within that set.Eligible non-winning commitments remain eligible, while open episodes may continue, interleave, suspend, or close.
  • 4.6 Behavior selection under constraint: The selector optimizes expected situated value over permitted candidates rather than choosing the best behavior overall.The mandate influences selection through intentions, constraints, and profile, while mandate-relative value preserves a separate comparison point.
  • 4.6 Behavior selection under constraint: The mandate sets a goal, scope, and risk-tolerance bar, making the least sufficient behavior the one that satisfies it without exceeding it.Going beyond what was warranted creates additional footprint without justification.
  • 4.6 Behavior selection under constraint: Candidate behavior includes action, communication, inquiry, delegation, monitoring, deferral, inhibition, and deliberate non-action.These are recurring behavior modes rather than a strict partition, and some may be realized through external actions.
  • 4.6 Behavior selection under constraint: Constraints are active governors: if an unconstrained optimum lies outside the permitted set, it must not be selected.The behavior selector maps the agentic state and profile to a policy-permitted behavior.
  • 4.6 Behavior selection under constraint: Behavior classes distinguish responsiveness to the world, self-initiation, social context, deliberation, reflection, adaptation, and invariant policy compliance.Adaptive change is directed by commitments and outcomes, whereas undirected change is drift.

4.7 Agenthood

Agenthood requires five organizing elements, while agency form is determined separately by whether an operative mandate exists. The framework distinguishes retained authority from goal origin and ownership, classifying agents as autonomous when no mandate remains.

  • The framework separates whether a system is an agent from which form its agency takes.Baseline elements determine agenthood, while the mandate determines form.
  • Minimal agenthood requires standing intent, situated perception, significance appraisal, constrained selection, and feedback with revision.The framework treats these as role-based elements rather than a component inventory; absence of any one means the system is not an agent in this sense.
  • A mandate records authorized goals, scope, duration, risk tolerance, and constraints, and remains revocable by the principal.Its behavioral role, rather than its record type, determines whether it functions as a mandate.
  • Autonomy means no operative mandate exists; goal origin, ownership, and construction history do not by themselves determine agency form.An agent may remain autonomous even when its goals were created or transferred by others, provided no authority is retained over their pursuit.
  • Retained and revocable authority classifies an agent as delegated, including when an owner can revoke or override its goals.Ownership concerns operation or shutdown, whereas a mandate concerns authority to act for a principal.

4.10 Delegated agency and its governance

Delegated agency is defined by an operative mandate that bounds behavior and keeps it answerable to a principal. Its governance framework contains authority, evaluates mandate-relative loss, records traces, and restricts adaptation within admissible updates.

  • Delegated agency exists when an operative mandate transfers bounded capacity to perceive, decide, and act while the principal retains authority.Governance becomes operative because behavior must remain answerable to the mandate.
  • Mandate-relative agency loss compares the mandate’s preferred behavior with the behavior selected by the agent.The comparison uses authority and evaluator-grounded representations, allowing errors in boundary translation or estimates of the principal’s condition to register.
  • Capability failure and agency loss are distinct: a capable agent can diverge from its mandate through boundary errors, incomplete information, poor coordination, or misdirected adaptation.The framework treats these sources as potentially overlapping classifications rather than asserting that their losses sum.
  • Authority is contained along delegation chains: delegated permission may shrink but cannot grow, and allowed behavior is the intersection of all binding hard boundaries.A single relevant non-negotiable boundary can forbid a behavior.
  • Bounded adaptation permits learning only within an admissible update relation and excludes authority, delegation scope, and hard constraints from ordinary learning.Provenance, versioning, and reversibility support correction and accountability where possible.

4.11 Symbiotic agency

Symbiotic agency is a continuing, coupled form of delegated agency designed to augment rather than replace the principal’s capabilities and judgment. Its evaluation combines form-specific agency elements with severity-weighted violations, while keeping capability and mandate alignment distinct.

  • Symbiotic agency requires a standing mandate, continuing coupling to the principal’s situation, and bounded personalization.These are necessary conditions beyond ordinary delegation.
  • Continuing coupling includes monitoring the principal’s context and using a calibrated estimate of the principal’s condition to condition behavior.The framework includes both exteroceptive context channels and allo-interoceptive channels.
  • Symbiotic behavior selection treats the principal’s capabilities and judgment as what the agent supports rather than replaces.Interruption cost and sensitivity make knowing when not to interrupt, expose, escalate, or interfere part of appraisal.
  • The capability score covers shared agentic elements and adds form-specific requirements for autonomous, delegated, or symbiotic agency.Shared elements include intent, perception, appraisal, activation, constrained selection, feedback, revision, and observable behavior records.
  • Severity-weighted alignment discounts capability as violations accumulate, while hard-constraint violations zero the score outright.The framework distinguishes graded lapses from inviolable-boundary violations and leaves severity calibration to the evaluation protocol.
  • High agency capability can coexist with high agency loss when proactive or adaptive behavior departs from the mandate.Capability, constraint satisfaction, mandate alignment, observability, and residual agency loss should therefore be evaluated separately.

5 Related Work and Conceptual Positioning

The paper positions governed proactive agency as an integration of commitment, appraisal, situated coupling, delegation, and evaluation around the activation problem. Its central additions are behavioral episodes, standing mandates, authority containment, and assessment beyond task completion.

  • Classical theories of agency: Prior theories contribute commitment, feedback, purposive organization, situated coupling, and performance evaluation, but do not decide when a standing commitment should become behavior.The paper identifies activation, user-relative restraint, and mandate sensitivity as complementary gaps across these traditions.
  • Computational appraisal and affective computing: Appraisal is extended for delegated agency through interruption cost and permission proximity, which evaluate attention claims and authority boundaries relative to a principal.These dimensions supplement standard appraisal counterparts because they explicitly encode whose cost and boundary are at stake.
  • Alternative formal accounts of agency: The framework preserves existing formal substrates while requiring separately represented world, user, and agent states, durable intent, significance, behavior readiness, mandates, boundaries, and restraint.It does not propose a replacement optimization calculus; these structures may be realized in POMDP or active-inference implementations.
  • Mixed-initiative interaction and interruption: Compared with mixed-initiative systems, the account broadens activation to standing mandates, richer behavior modes, and governance across continuing delegation.The agent may ask, monitor, remember, wait, escalate, inhibit, or deliberately refrain, not merely act or defer.
  • Delegated and multi-agent agency: A mandate anchors delegated and multi-agent behavior by preserving authority, scope, consent, accountability, governed traces, bounded adaptation, and mandate-relative agency loss.This formalizes coordination failure and information asymmetry against authorization rather than task success alone.
  • Comparative summary: The paper treats behavioral episodes—not isolated actions—as the unit of analysis and organizes them through significance-gated activation that includes inquiry, monitoring, deferral, inhibition, and restraint.This integrated structure connects component traditions while addressing whether behavior should occur, when, and in what mode.

6 Agency Signatures

Section 6 classifies systems by their agency signatures: the structures they instantiate and how fully they realize them. It separates agenthood from agency form, then uses coverage and placement to distinguish autonomous, delegated, and symbiotic agency.

  • Classification: An agency signature records which structures an approach or system instantiates, rather than assigning one undifferentiated agency level.The classification treats agency as layered and asks which structures are present.
  • Structure: The shared baseline includes agenthood structures, while autonomous, delegated, and symbiotic groups add progressively specific authority, governance, coupling, and personalization structures.The groups are cumulative, with symbiotic agency extending delegated agency.
  • Scoring: Element scores are 1 for first-class structures, 1/2 for partial or implicit structures, and 0 when absent.Scores reflect whether an element has its own representation or mechanism and participates in behavior selection.
  • Placement: Placement first tests the agenthood minimum, then selects autonomous, delegated, or symbiotic form from mandate and symbiotic conditions.Coverage can change without changing placement, whereas a single missing defining element can determine placement.
  • Placement: Agenthood requires standing intent, situated perception, significance appraisal, constrained selection, and feedback with revision; missing any one prevents agent placement.A system with an authorization record but without the agenthood minimum may still be described as delegated, but not as a delegated agent.
  • Comparison: The comparison finds that no prior approach combines appraisal-triggered activation with behavior answerable to an operative mandate or covers all delegated and symbiotic elements together.The result concerns the compared approaches, not a single implementation benchmark.

7 Measuring Agency and Agency Loss

Section 7 turns the theory into a multidimensional evaluation framework rather than a single agency score. It measures realization, authorization, constraint adherence, activation quality, timing, calibration, provenance, coordination, and adaptation across system types.

  • Evaluation framework: The framework evaluates how instantiated agency structures operate, because systems can activate at the wrong time, exceed mandates, obscure decisions, or lose agency after governance.The paper explicitly rejects collapsing agency into a single task score.
  • Core measures: Element realization estimates each agency element’s operational degree, preferably from behavioral records when the element leaves a footprint.Audit-based scores remain necessary for elements such as consent-gated perception and adaptation bounds.
  • Governance: Mandate alignment, constraint-violation severity, authority integrity, and provenance measure whether delegated behavior remains authorized and reconstructable.These measures cover agreement with the mandate, permitted-set violations, delegation containment, and trace coverage.
  • Activation: Inappropriate-intervention, missed-intervention, and timing measures capture whether the agent acts, refrains, or delays appropriately.Timing includes interruption cost given the principal’s inferred condition and delay cost from deferral.
  • Calibration and adaptation: Principal-state calibration measures estimation error and correction after feedback, while coordination and adaptation measures address shared-state failures and bounded updates.Adaptation integrity includes profile drift, unauthorized updates, reversibility, and adherence to bounded adaptation.
  • Illustrative cases: The three illustrative cases span measures that are simple or inapplicable for a thermostat, log-estimable for a coding agent, and broadly applicable to a symbiotic assistant.The symbiotic target additionally requires evidence often absent from current records, including gate decisions, traceable restraint, and principal judgments.
  • Benchmarks: The proposed scenario families define an evaluation agenda, but numerical results are deferred until scenarios, baselines, data, and protocols are constructed and validated.The paper presents estimators and future benchmark requirements rather than empirical scores.

8 Design Implications and a Reference Layer Stack

Section 8 translates the formal model into implementation requirements and a reference layer stack. It permits different realizations while requiring inspectable governance boundaries, including consent, authority containment, traceability, reversible adaptation, and external shutdown.

  • Requirements: The implementation requirements apply across an orchestrated runtime or integrated model, with some requirements depending on delegated or symbiotic agency.The requirements preserve responsibilities and governance boundaries rather than mandate one component layout.
  • Perception and inference: Perception must be partitioned and policy-gated, while principal-state inference remains distributional, uncertainty-carrying, and corrigible.The inferred principal condition is not treated as direct access to the principal’s internal state.
  • Activation and restraint: Durable intent is separated from activation, and the behavior vocabulary includes logged restraint actions such as waiting, monitoring, deferring, and inhibiting.Activation is computed through appraisal rather than triggered solely by prompt arrival; doing nothing becomes a recorded decision.
  • Governance: Governance requires non-learned consent and execution enforcement, versioned mandates with provenance and revocation, authority containment, severity-monitored violations, and bounded reversible adaptation.The action firewall and consent gate remain outside the learned system and cannot be bypassed by the agency module.
  • Recovery and accountability: Graceful degradation narrows authority and preserves evidence, while reconstructable episodes record activation, restraint, authority, evidence, and outcomes.Every deployment also retains identifiable ownership with externally enforced suspension or shutdown.
  • Realization strategies: An orchestrated runtime uses explicit services around a language-model cognition core, making interfaces, constraints, permissions, and traces easier to inspect.The strategy fits current engineering practice in which models operate inside coordinating runtimes.
  • Realization strategies: ACAM proposes a stateful multimodal model with continuous perception-to-behavior processing, exposed readouts, a unified behavior vocabulary, and an external governance shell.Whether joint multimodal appraisal outperforms late joining remains an empirical question, and ACAM’s training and evaluation are outside this paper’s scope.

9 Open Problems

Section 9 identifies open problems for governed proactive agency, delegated authority, and symbiotic coupling. The challenges concern robustness, evidence, privacy, authority, personalization, scoring reliability, and conflicts involving principals or third parties.

  • Scope: The open problems are tiered: some concern proactive agency generally, others arise from delegation, and others specifically require continuing coupling to a principal.This organization follows the theory’s agency taxonomy.
  • Robustness: Gate manipulation and adversarial significance inflation could lower activation barriers or distort perceived urgency, motivating robust appraisal architectures and evaluation.The emergency override is identified as an attack surface, while adversarial content can make appraisal vulnerable to prompt injection.
  • Evidence: Significance-label acquisition remains difficult because judgments about what mattered, for whom, and when are subjective, contextual, and expensive.The paper leaves open which supervision signals and fidelity levels are sufficient.
  • Evaluation: Counterfactual restraint is hard to score because correct non-interventions often leave no observed outcome.The paper proposes scenario suites as one route beyond this limitation.
  • Delegation: Delegation raises unresolved questions about legitimate multi-principal conflict rules, path-wide containment under graph changes, and mandate translation loss.These problems concern auditing disagreements, runtime invariants, and differences between principal intent and internalized policy or profile.
  • Adaptation: Personalization introduces profile poisoning and gradual drift, requiring detection, attribution, and reversal mechanisms that preserve bounded adaptation.Feedback itself can be manipulated, making drift difficult to detect through review alone.
  • Interpretability: Inspectable appraisal readouts may rationalize decisions after the fact rather than faithfully reveal the computation that produced them.The open question is when readouts count as evidence rather than rationalization.
  • Privacy: Symbiotic sensing must reconcile intimate allo-interoceptive signals with on-device processing, data minimization, consent granularity, and retention policy.The challenge is to make principal-state sensing governable and defensible.

10 Ethical and Societal Considerations

The paper treats proactive agency as an ethical and societal governance problem spanning perception, inference, delegation, ownership, records, and deployment at scale. It requires consent, calibration, revocability, answerability, and limits on surveillance and attention capture.

  • Perception and inference: Perception is a primary ethical boundary because expanded sensing can increase helpfulness while also increasing intrusion, misinterpretation, manipulation, and surveillance risks.The paper requires consent, locality, minimization, explanation, and user control, with a non-learned consent gate that the system cannot relax through learning.
  • Perception and inference: Inferences about stress, fatigue, attention, or emotion are fallible constructions that must remain uncertain, correctable, and answerable to the principal.Treating such estimates as facts risks misreading the principal, manipulation, or paternalistic intervention.
  • Delegation and accountability: Delegated agency must remain attributable, revocable, and answerable because standing mandates and adaptive profiles create durable capacity to act for someone else.Authority containment and external enforcement are required when self-initiated behavior escapes policy.
  • Delegation and accountability: Governance records support accountability by capturing authority, evidence, decisions, outcomes, restraint, and sub-threshold evaluations, but they also contain sensitive inferred information.Consent, minimization, retention limits, and scoped audit rights are needed to prevent accountability from becoming surveillance by another party.
  • Societal scale: At societal scale, unequal access to governed systems and competition among proactive agents for human attention create distributional and coordination risks.The paper argues that governance should not become a premium feature and that cross-agent restraint and limits on attention capture are needed.

11 Conclusion

The conclusion presents governed proactive agency as an integrated account in which perception, intent, appraisal, constraint, and feedback organize behavior across time. It turns proactivity into something that can be specified, measured, traced, and evaluated while preserving authorization, calibration, restraint, and answerability.

  • Conclusion: Proactive agency concerns whether behavior should occur, when it should occur, and in what mode, making restraint a selected behavior rather than a failure to act.The activation problem distinguishes proactive agency from task execution.
  • Conclusion: The taxonomy distinguishes autonomous, delegated, and symbiotic agency, with symbiotic agency defined as delegated agency under a standing mandate and continuing coupling to the principal.Its purpose is to augment the principal’s capabilities without replacing their judgment.
  • Conclusion: The paper’s main contribution is an integrated structure connecting appraised significance with mandate-answerable behavior and continuing symbiotic coupling.It defines behavioral episodes, perception, allo-interoception, intent, appraisal, readiness, activation, selection, mandates, authority containment, bounded adaptation, and mandate-relative agency loss.
  • Conclusion: Agency signatures, evaluation measures, implementation requirements, and a reference layer stack enable comparison by warranted, constrained, traceable, and answerable behavior rather than task success alone.The evaluation framework includes activation quality, timing, constraint violations, mandate alignment, observability, calibration, delegation integrity, and adaptation integrity.
  • Conclusion: Future work includes formalizing and evaluating the Affective-Conative Agency Model, realizing and releasing the seven-family benchmark, and further developing the theory’s implications.The paper identifies restraint, governance, calibration, and operational robustness as subjects for continued formal development.

A Detailed comparison of neighboring frameworks

The detailed comparison tables organize neighboring approaches by baseline agency elements and by elements of autonomous, delegated, and symbiotic forms. Their entries describe typical formulations, with limited or partial representation distinguished from first-class treatment.

  • Interpretation of entries: The tables treat Limited and Partial as representations that are not normally distinct first-class objects, corresponding to the half-filled symbol in Table 4.The entries are typical formulations rather than an exhaustive account of what an approach could express after extension.
  • Baseline elements: Table 8 compares neighboring approaches element by element across the baseline elements of agency.It reports the typical formulation of each element in each approach.
  • Three agency forms: Table 9 compares neighboring approaches element by element across the elements of the autonomous, delegated, and symbiotic forms.The comparison uses the form-specific column groups described for the agency taxonomy.
Loading 2609.07741v1…